Safeguard
Supply Chain Security•

How Lasers Are Used to Hack Devices and How to Defend

Laser attacks on devices explained: Light Commands, LiDAR spoofing, laser fault injection on chips, LED leaks and camera blinding, and how to defend.

Safeguard Security Team
Updated 63 min read

Lasers and other light sources can attack devices in a handful of documented ways. Light aimed at a microphone can be turned into fake voice commands. Light timed against a LiDAR can add obstacles that are not there or hide ones that are. Light focused on a bare chip can flip bits and skip security checks, which is how evaluation labs test smart cards and secure elements. Light can also leak out: a power LED that flickers with a device's workload can give away sound or key material to a camera. Almost every case comes from published research, the physical effect is usually hard to remove, and the most effective defences sit in the firmware and software that decide what to do with the signal.

That is the short answer. This guide is the long one. It explains why sensors and chips respond to light at all, and walks through the research that defines this field: Light Commands, the 2019 work that injected voice commands into smart speakers with a laser; a decade of LiDAR spoofing studies against self-driving perception; laser fault injection on chips from a 2002 camera flash experiment to a 2026 hardware wallet card; optical side channels from blinking modem LEDs to video of a card reader's power light; infrared tricks against cameras; the security of laser links between buildings and between satellites; and, briefly, laser pointers and aircraft. It then covers how chip makers test and certify against these attacks, how to detect them, hardening checklists for home users, businesses and manufacturers, and the software supply chain inside every one of these devices.

It is written from the attacker's point of view for defenders. For each attack you will find how it works and why, what the attacker needs, what the published results showed, the impact, how to detect it and how to defend against it. You will not find exploit code, equipment settings, tool command lines or step-by-step instructions. This post covers civil and commercial research only. Where a tool, a lab or a contest is named, it is named as context for authorised testing.

A note on the law and on safety. Testing devices you do not own, or have no written permission to test, is illegal in most countries. Shining a laser at aircraft, vehicles or people is dangerous and is an offence in many places. Lasers above the power of a common pointer can injure eyes; the Light Commands researchers themselves ask anyone repeating their work to get formal laser safety training first. Everything in this guide is for protecting your own devices, your organisation's systems with authorisation, or the products you build.

This post is part of a device security series. The companion guides on how smart home devices get hacked, how cars and EV chargers get hacked and how laptops get hacked cover the devices where these attacks land. Our post on how computer chips are made, from sand to software explains where the secure chips discussed below, and the software that runs on them, come from.

Contents

  1. What is a light-based attack on a device?
  2. Why do sensors and chips respond to light at all?
  3. What is the attack surface, and what does an attacker need?
  4. How did Light Commands inject voice commands with a laser?
  5. How is LiDAR spoofed, and can it fool a self-driving car?
  6. What is laser fault injection, and how do labs use it on chips?
  7. How do hardware makers test and certify against these attacks?
  8. How can an LED leak data or sound?
  9. Can infrared light blind or fool a security camera?
  10. How secure are laser links between buildings and satellites?
  11. Why are laser pointers a safety problem for aviation?
  12. How should you weigh these risks?
  13. How do you detect light-based attacks?
  14. Why is this a firmware and software supply chain problem?
  15. How should home users protect themselves?
  16. How should businesses and operators protect their systems?
  17. How should manufacturers build devices that resist light-based attacks?
  18. How Safeguard helps teams that build these devices
  19. The takeaway
  20. Frequently asked questions
  21. Sources

What is a light-based attack on a device?

A light-based attack uses light, often from a laser, to make a device's sensor or circuit produce a signal it should not, or to read information a device gives off as light. The light itself does nothing clever. The damage comes from firmware that trusts whatever the sensor reports and acts on it.

A few terms first, because the rest of the guide uses them.

A laser produces a narrow, intense beam of light at one wavelength. Its value to an attacker is reach and precision: it can put a lot of light on a very small spot from far away. A modulated beam is one whose brightness changes over time in a pattern. Change the brightness the way a sound wave changes air pressure, and the beam carries an audio signal.

A sensor turns something physical into an electrical signal. A MEMS microphone (micro-electro-mechanical system) is the tiny microphone chip inside phones and smart speakers, with a thin diaphragm that moves with sound. A LiDAR (light detection and ranging) sends pulses of laser light and times the echoes to measure distance. A camera's image sensor is a grid of light-sensitive cells.

A side channel is information a device leaks without meaning to, through something other than its normal outputs: power use, timing, radio emissions, sound or light. A fault injection attack disturbs a chip while it runs, through voltage, clock, electromagnetic pulses or light, so that it makes a mistake an attacker can use. Laser fault injection (LFI) is the light-based version.

Put together, the documented attacks fall into four groups, shown in Figure 1. Light can go in to a sensor or a chip. Light can come out of a device and give away what it is doing. Light can carry data between buildings and satellites. And light can be aimed at people operating vehicles.

Map of seven ways light becomes an attack path: light going into MEMS microphones, LiDAR sensors, cameras and chip dies; light coming out of status and power LEDs; light as a link in free space optical systems; and light aimed at pilots and drivers

Figure 1. Seven documented ways light becomes an attack path, grouped by direction.

Why do sensors and chips respond to light at all?

Silicon reacts to light. When enough light reaches a transistor or a sensor chip it creates a small current, and light absorbed by a surface can heat it and make it move. Sensors are built to turn tiny changes into signals, so they turn unwanted light into signals too.

Three physical effects explain almost everything in this guide.

The first is the photoelectric effect in silicon. Light of the right energy frees electrons in a semiconductor and creates a current. Image sensors and photodiodes are designed around this. Other chips are not, but they are still made of silicon. In Skorobogatov and Anderson's words from their 2002 paper on optical fault induction, "Illumination of a target transistor causes it to conduct, thereby inducing a transient fault." That is the basis of laser fault injection.

The second is the photoacoustic effect. Light absorbed by a thin surface heats it, the heat makes it expand and contract, and a modulated beam makes it vibrate. The Light Commands researchers suspect this plays a part in why a laser can move a microphone's diaphragm, alongside a photocurrent in the microphone's own chip.

The third is light as a measurement. LiDAR and cameras exist to measure light, and they cannot tell, from the photons alone, whether light came from their own emitter, from the sun or from someone else. A LiDAR that receives a pulse at the right moment records an echo. A camera that receives a bright beam adjusts its exposure. Neither is malfunctioning. They are doing their job with false input.

The pattern repeats in the other direction. Electronics draw more or less current depending on what they are doing, and a status LED wired to that supply gets slightly brighter or dimmer. Human eyes cannot see the flicker. A photodiode or a fast camera can.

None of these effects can be patched out of physics. What can change is what the device does next: whether it asks for confirmation, compares several sensors, checks that a reading is physically plausible, detects the disturbance and stops, or keeps its secrets in hardware built to resist it. That is why the rest of this guide keeps returning to firmware and software.

What is the attack surface, and what does an attacker need?

The attack surface is every place light reaches a component that turns it into a decision: microphones, LiDARs, cameras, bare chips and optical links, plus LEDs that let light out. Most light attacks need line of sight. Chip attacks need the device in hand and a specialist lab.

The table below lists what each documented attack class has needed. It describes requirements, not methods.

Attack classTargetWhat the attacker needsTypical setting in the research
Light injection into microphonesSmart speakers, phones, tabletsLine of sight to the microphone port, a bright modulated light source, aiming optics for long rangeLab corridors and through a window across buildings
LiDAR spoofing and removalSelf-driving and driver assistance perceptionLine of sight to the sensor, precise timing against its pulses, knowledge of the sensor modelLabs, controlled tests and simulators
Camera glare and infrared tricksSecurity cameras, car cameras, face recognitionLine of sight, bright or infrared light, sometimes a model of the recognition systemLabs
Laser fault injectionSmart cards, secure elements, microcontrollersThe chip itself, an opened package, a laser station, weeks of expertiseEvaluation and research labs
Optical side channelsDevices with LEDs wired to their power supplyA view of the LED, a photodiode or camera, signal analysisLabs, at distances of metres to tens of metres
Optical link interception or interferenceBuilding and satellite laser linksAccess to the beam path, or to the terminals at each endTheory and physics; terminals are the practical target
Laser illumination of pilotsAircraftA handheld laser and line of sightReal world, reported in large numbers

Two observations shape the defences. First, nearly all of the injection attacks need line of sight, so placement, covers and shielding work. Second, the attacks that need the least equipment, aiming a beam at a sensor or a person, usually produce disruption rather than control. The ones that give real control, such as faulting a chip to skip a password check, need the device in hand and serious resources.

Timeline of published light-based attack research from 2002 to 2026, from LED leaks and camera flash faults to LiDAR spoofing, Light Commands, Glowworm, video-based cryptanalysis and a 2026 laser fault on a hardware wallet card

Figure 2. Twenty four years of published research on light-based attacks against civil devices.

How did Light Commands inject voice commands with a laser?

Light Commands showed that a laser whose brightness follows a voice recording, aimed at a smart speaker's microphone port, makes the microphone report that voice. Google, Amazon, Apple and Facebook voice devices accepted the commands, in one case from 110 metres and through a window.

What the research did

Light Commands was published in 2019 by Takeshi Sugawara of the University of Electro-Communications in Tokyo and Benjamin Cyr, Sara Rampazzi, Daniel Genkin and Kevin Fu of the University of Michigan, and presented at USENIX Security 2020. The paper describes "a new class of signal injection attacks on microphones by physically converting light to sound." The team aimed amplitude-modulated light, light whose brightness rises and falls like a sound wave, at the opening above a device's MEMS microphone, and the device heard a voice command.

Why it works

A MEMS microphone has a diaphragm and a small chip, an ASIC, that turns the diaphragm's movement into an electrical signal. According to The Register's report on the paper, the researchers explain that strong light on the chip "induces a photocurrent across a transistor," and they suspect that light-driven heating also moves the diaphragm. Either way, the microphone "treat[s] light as sound." The project site adds that the effect is generally independent of the laser's colour, and that any sufficiently bright light may work.

The voice assistant then does what it always does: it listens for a wake word and a command. It has no way to know the "voice" arrived as light.

Six-stage diagram of Light Commands from a modulated light source through line of sight, the microphone port, the MEMS microphone and the voice assistant to an action, with a defence beneath each stage

Figure 3. Light Commands, conceptually, with a defence at every stage.

What it needed

The project site lists off-the-shelf parts: laser pointers, a laser driver, an audio amplifier and, for long range, a telephoto lens. The Register reported that the researchers' own setup, with a scientific-grade laser driver, cost about $1,500, and that they believed cheaper parts could work. The attacker needs a clear line of sight to the microphone port and steady aim, which is easier against a speaker that sits still on a shelf than against a phone in someone's hand.

What the results showed

The project site's table covers 17 devices. A few figures from it:

  • A Google Home responded to as little as 0.5 mW of laser power at 30 cm, and worked at more than 110 metres with 5 mW, the longest corridor the team had.
  • Echo Plus units (first and second generation) and the EchoBee 4 thermostat with Alexa also worked at long range with 5 mW.
  • Phones and tablets needed far more power and shorter range. The iPhone XR needed 21 mW and worked at 10 metres at 60 mW; the Galaxy S9 and Pixel 2 needed 60 mW and 46 mW and worked at about 5 metres.

The site notes that the list is not exhaustive and that "any system that uses MEMS microphones and acts on this data without additional user confirmation might be vulnerable."

The impact

What a light command can do depends entirely on what the voice assistant controls. The researchers showed they could operate smart switches, open garage doors, make online purchases, remotely unlock and start certain vehicles linked to an account, and open smart locks by brute-forcing a voice PIN. The project site also points out that speaker recognition was off by default on the smart speakers they tested, and that where it was on, it checked the wake word rather than the whole command. The researchers said they had seen no sign that the technique had been used maliciously.

How to detect it

The attack is silent, but not invisible. The project site notes that a user may see the laser's reflection on the device, and may notice the assistant's spoken replies or its status lights responding when nobody spoke. The most reliable signal is the assistant's activity history: commands at odd times, or commands nobody in the house remembers giving. Smart home platforms that log every lock, garage door and purchase event make this easier to spot.

How to defend against it

  • Ask before acting. The fix that matters most is software. Requiring a PIN, a voice match on the full command, or a confirmation on the owner's phone before unlocking, opening, buying or starting anything removes most of the impact. The researchers also suggest asking a randomised question before executing a sensitive command.
  • Fuse microphones. Smart speakers have several microphones. A single beam usually reaches only one. Firmware that compares them and rejects a command heard by only one microphone blocks the simple form of the attack.
  • Block the light path. An opaque cover or barrier over the microphone port helps. The Register noted that dense covers helped and that a light covering of fabric or foam was not enough on its own for some devices. The researchers warn that more laser power can overcome a weak barrier.
  • Place devices sensibly. A speaker that cannot be seen from outside a window cannot be hit from outside a window.
  • Limit what voice can do. Do not link door locks, garage doors or car functions to a voice assistant unless the action needs a second factor.

The Register reports that the researchers informed Apple, Google and Facebook so they could design mitigations. For makers, the lesson goes beyond microphones: a sensor's output is input, and input that triggers a physical or financial action deserves confirmation. Our guide to how smart home devices get hacked covers the rest of the voice assistant and smart lock attack surface.

How is LiDAR spoofed, and can it fool a self-driving car?

A LiDAR measures distance by timing laser echoes. Light from someone else, timed to arrive like an echo, can add points for an obstacle that is not there or drown out real echoes so an obstacle disappears. Researchers have shown both in labs and simulators. Sensor fusion, randomised timing and plausibility checks push back.

How LiDAR works

A LiDAR fires short laser pulses, often in the near infrared, and measures how long each takes to come back. Distance is the round-trip time multiplied by the speed of light, divided by two. A rotating or scanning LiDAR does this many thousands of times per second in many directions and builds a point cloud, a 3D picture of the surroundings. Perception software, often machine learning models, turns the point cloud into a list of objects: cars, pedestrians, cyclists, barriers. Planning software then decides whether to brake, steer or carry on.

The weak point is simple. A LiDAR knows when it fired. It does not inherently know whether the light that came back is its own.

Diagram of LiDAR ranging, phantom point injection and real point removal, with published results and six defences: randomised timing, a shorter listening window, probing more than once, physical consistency checks, sensor fusion and safe fallback

Figure 4. How LiDAR is fooled, conceptually, and what pushes back.

What the research showed

The published work builds up over a decade.

2015: commodity parts. Jonathan Petit, Bas Stottelaar, Michael Feiri and Frank Kargl presented remote attacks on a camera and a LiDAR at Black Hat Europe, deliberately limiting themselves to "commodity hardware (below 60 US$)." Against an ibeo LUX 3 LiDAR they showed that its pulses were not tied to the sensor that emitted them, so recorded pulses could be relayed and replayed. During their relay test, a LiDAR facing a wall one metre away reported echoes from objects at 20 and 50 metres. They concluded that "a cheap transceiver was able to inject fake objects that are successfully detected and tracked."

2019: fooling the perception software. Yulong Cao and colleagues at the University of Michigan and UC Irvine ran the first security study of LiDAR-based perception (ACM CCS 2019) against Baidu Apollo, an open source autonomous driving system. They found that simply spoofing points was not enough to fool the machine learning model, so they treated the attack as an optimisation problem. A direct approach placed a fake obstacle just in front of the vehicle only about 30 percent of the time; combining global sampling with optimisation raised that to around 75 percent. In Apollo's simulator, using real driving traces released by the Apollo team, they showed two consequences: an emergency brake attack, which could cause a rear-end collision, and an "AV freezing" attack, which could leave a vehicle waiting at a red light stuck in the intersection. The authors also proposed defences at the system, sensor and model levels.

2020: a general attack, and a defence. Jiachen Sun, Yulong Cao, Qi Alfred Chen and Z. Morley Mao found that perception models ignored occlusion patterns in point clouds, the shadows that real objects cast in LiDAR data. A black-box spoofing attack built on that gap reached about 80 percent mean success across the target models. Their defence, CARLO, treats occlusion as a physical invariant and checks it, cutting mean attack success to 5.5 percent. A model architecture that builds the same physical features into learning, called SVF, cut it further to about 2.3 percent.

2023: making obstacles disappear. In You Can't See Me (USENIX Security 2023), Cao, Bhupathiraju, Naghavi, Rampazzi, Mao and Sugawara showed the opposite direction. Physical Removal Attacks use laser spoofing to selectively remove a real obstacle's points at the sensor level, before perception sees them. Tested against the Apollo, Autoware and PointPillars detectors and evaluated in the LGSVL simulator, the attack removed 90 percent of a target obstacle's points in 92.7 percent of moving-vehicle scenarios. It also defeated two existing defences, and the authors proposed two enhanced ones.

2024: newer sensors. Takami Sato and colleagues ran the first large-scale measurement study (NDSS 2024) across nine popular LiDARs, including new-generation models, and three types of object detectors trained on five datasets. They reported 15 new results, several of which challenged earlier assumptions about what spoofing can do, identified a removal attack that works on new-generation LiDARs, and discussed defences. The practical message for engineers: results on one sensor model do not transfer automatically, in either direction.

The impact, kept in proportion

A phantom obstacle can cause an unnecessary emergency stop, which risks a rear-end collision. A removed obstacle is the more serious case, because the car may not brake for something real. Both were shown in labs, controlled experiments and simulators. Automated vehicles do not rely on one sensor. As Petit and colleagues note, they carry "a multitude of sensors, using different physical properties," so an attacker who fools one still has to get past the others and past the safety logic that compares them. The research matters because it tells engineers where those layers need to hold.

Light also works in the other direction. In 2025, The Drive reported that Volvo's support guidance for the EX90 says "Do not point a camera directly at the lidar," and a Volvo spokesperson said that "the laser light emitted by the lidar can potentially damage the camera's sensor or affect its performance." That is a safety note rather than an attack, but it is a reminder that eye-safe light can still harm another sensor.

How to detect it

Detection belongs in the vehicle's software. Useful signals include points that change between repeated probes of the same direction, objects that appear without the occlusion shadow a real object would cast, objects that one sensor sees and the others do not, and returns arriving at times the sensor did not expect. Fleets should log sensor disagreements and safety interventions centrally, so a pattern at one location or time stands out.

How to defend against it

  • Randomise or skip pulses. Petit and colleagues suggested varying the pulse interval or skipping pulses unpredictably so outside light cannot synchronise. A LiDAR that hears a reply to a pulse it never sent has caught an attack.
  • Probe more than once. Measuring the same direction several times and rejecting inconsistent distances defeats unsynchronised interference, at some cost to scan rate.
  • Shorten the listening window. A shorter window between pulses leaves less time for a fake echo, at the cost of maximum range.
  • Check physical consistency. CARLO's result shows how much a plausibility check on occlusion and shape can cut attack success.
  • Fuse sensors. Camera, radar and LiDAR rely on different physics. Making them agree raises the bar.
  • Fail safe. When sensors disagree, a vehicle should slow, stop safely or hand over, not guess.

Most of these defences live in LiDAR firmware or perception software, which makes them software supply chain items like any other. Our guide to how cars and EV chargers get hacked covers the rest of the vehicle software stack.

What is laser fault injection, and how do labs use it on chips?

Laser fault injection fires precisely timed laser pulses at an opened chip so that transistors briefly conduct, flipping a bit or skipping an instruction. Evaluation labs use it to test smart cards and secure elements. It needs the chip in hand, specialist equipment and expertise, and chips resist it with shields, light sensors, redundancy and hardened firmware.

Why chips that hold keys are tested this way

Smart cards, payment chips, SIM cards, secure elements in phones and hardware wallets, and security chips in IoT devices exist to keep a secret inside even when an attacker holds the device. A firmware check such as "is this PIN correct?" or "is this update signed?" ends in a single decision. If a fault can make the chip take the wrong branch once, at the right moment, the check is bypassed. Fault injection is how attackers try to cause that, and how evaluators prove a chip resists it.

Where it started

In 2002, Sergei Skorobogatov and Ross Anderson of the University of Cambridge published Optical Fault Induction Attacks. With a camera flashgun bought second-hand for about $30, mounted on a microscope, they could set or reset individual bits of SRAM in an opened PIC16F84 microcontroller. They wrote that the attack "does not even require expensive laser equipment" and warned that optical probing could "induce errors in cryptographic computations or protocols, and to disrupt the processor's control flow." They also noted that top-layer metal shielding and bus encryption "may make an attack using these techniques more complicated, but are not enough," because a determined attacker can come through the back of the chip with an infrared laser. Their proposed defence was self-timed dual-rail logic, where every bit travels on two wires and an impossible combination raises an alarm that locks the chip.

What a modern laser station is

Today's equipment is built for exactly this testing. Keysight's description of its fault injection laser systems explains that LFI "uses highly controlled laser pulses to temporarily disturb the operation of a semiconductor device," and that its users are evaluation laboratories, semiconductor companies, certification facilities and product security teams. A front-side attack goes through the top of the chip, where metal layers and circuitry are; a back-side attack goes through the silicon substrate, which is "often used when metal layers on the front side obstruct access." Either way, the package usually has to be opened first. Stations combine a microscope, one or two lasers, a motorised stage and timing control, so that a pulse lands on a chosen spot at a chosen moment.

Diagram of what a laser fault injection evaluation needs, from the chip in hand to an opened package, a laser station, time and expertise, next to six layers of defence: active metal shield, light and glitch sensors, redundant logic, protected memory, hardened firmware and an update path

Figure 5. Laser fault injection needs the chip in hand and a lab. Secure chips resist it in layers.

What published lab work has shown

A secure memory, 2020. Olivier Hériveaux of Ledger's Donjon security team presented black-box laser fault injection on the Microchip ATECC508A at SSTIC 2020. The chip, used in the Coldcard Mk2 hardware wallet to store the wallet's seed, had a top-metal shield against probing, an internal clock and an internally regulated core voltage against glitches. The paper notes there was "no mention to resistance against laser fault injection (such as light detectors for instance)" in its documentation, which is why the team tested it. They identified a way to read protected data slots in less than a month, destroyed two chips along the way, and once the parameters were known extracted test data from a fresh chip "in less than 2 minutes." The paper also points out that the successor chip, the ATECC608A, added more software security hardening.

A hardware wallet card, 2026. In July 2026, the Donjon team published an attack on Tangem cards, which are built on a Samsung S3D232A secure element that the post says is certified to EAL6+. A laser fault on a single conditional check in the card's password-change code let them set a new password without the old one or the backup card. They report that the attack needs physical access, specialised equipment costing around $250,000, significant expertise and extensive characterisation, and that "the attack is physical and invasive so it can't be done covertly." They disclosed it to Tangem on 10 February 2026 and state that it "cannot be patched," because "the cards have no firmware update mechanism." Tangem's response, reported by The Block, called the risk to everyday users "virtually non-existent," pointed to the lab equipment and physical possession required, and noted that Ledger is a competitor. The researchers' own summary agrees on the practical risk: "The only real risk is a lost or stolen card."

Both cases are worth reading for their recommendations rather than their drama. The Donjon post recommends verifying an authorisation decision at several points in the code, so an attacker would need several simultaneous faults, and encoding security-critical flags with values far apart in binary, a large Hamming distance, so that a single flipped bit is detected rather than accepted. Its conclusion is the one every chip and firmware team should remember: "EAL6+ certification does not guarantee immunity to fault injection."

How chips resist laser faults

No single countermeasure is enough, so secure chips stack them. Figure 5 shows the layers.

  • Active shields. A fine mesh of wires over the circuit, monitored by the chip, detects cutting and probing. It also blocks a front-side laser's view of the logic beneath.
  • Light and glitch sensors. Detectors for light, voltage, clock and temperature anomalies spread across the die, wired to reset or lock the chip. Some chips may also keep a persistent count of detected faults and disable themselves after too many; the Donjon team suspected such a counter on the Tangem card's chip and had to work around it.
  • Redundancy. Dual-rail logic, duplicated computation and compare-before-output make a single fault produce an error rather than a wrong answer.
  • Protected memory. Memory encryption and error-detecting codes make it harder to read or silently alter stored secrets.
  • Hardened firmware. Multiple checks, safe flag encodings, random delays and verification of results before release mean a single skipped instruction is not enough.
  • An update path. The Tangem case shows the cost of having none. If the weakness is in firmware, a signed update can fix it. If the card cannot be updated, it cannot be fixed.

Impact and detection

The impact of a successful fault is the loss of one device's secrets: a key, a PIN-protected function, a wallet's seed. The attack scales poorly, because each device has to be opened and characterised. Detection is mostly physical. Laser attacks are invasive and usually leave a chip with an opened, often damaged package, and chips with fault counters may lock themselves. For owners, the practical defence is the oldest one: keep hardware wallets, payment cards and key cards in your possession, and treat a lost one as compromised.

How do hardware makers test and certify against these attacks?

Chips that protect keys and payments go to accredited labs that attack them, including with lasers, under schemes such as Common Criteria, EMVCo, SESIP and PSA Certified. Certification raises the cost of an attack and gives buyers evidence, but it does not prove a chip is immune.

Common Criteria

Common Criteria is the international standard for security evaluation, published as ISO/IEC 15408. A protection profile sets out security requirements for a class of products, such as smart cards. A vendor's security target states the security properties of its specific product, and each product is evaluated against its own target. Evaluation assurance levels, from EAL1 to EAL7, describe how rigorous the evaluation was, and higher levels do not automatically mean "better security." Testing labs must comply with ISO/IEC 17025. For smart card chips, these evaluations include physical attack testing, which is where laser stations come in; EMVCo, below, accepts Common Criteria results for the chip stage of payment certification.

EMVCo

Payment chips and cards are evaluated under EMVCo's process. Keysight's description of that process splits it into three stages: the IC evaluation covers the chip hardware and its on-chip crypto libraries, the platform evaluation covers the card operating system and builds on the IC results, and the ICC evaluation covers the complete product, platform and applications. Penetration testing targets "high attack potential," described as "a well-resourced, well-informed adversary with access to specialist equipment," and includes "voltage, clock, electromagnetic, and laser fault injection" alongside power and electromagnetic side channel analysis. EMVCo recognises Common Criteria results in its IC certification process, so many chip vendors run one evaluation for both.

SESIP and PSA Certified for IoT chips

Connected devices use lighter schemes built on the same ideas. SESIP, GlobalPlatform's Security Evaluation Standard for IoT Platforms, is "a methodology that reduces the cost, complexity and effort of security evaluation and certification" by reusing certified components. It has been adopted as the European Standard EN 17927 and defines five assurance levels. PSA Certified Level 3 evaluates a chip's root of trust and "includes more attack types, more sophisticated side channel and perturbation attacks as well as physical attacks," with 35 days of white-box evaluation under either a SESIP or a CSPN-style profile.

Four evaluation schemes compared: Common Criteria, EMVCo, SESIP and PSA Certified Level 3, with the lesson that certification raises the cost of an attack but does not make a chip immune

Figure 6. How chips are evaluated against physical attacks, including laser fault injection.

How the labs work

The process is broadly the same everywhere. The lab studies the design documentation to find likely weak points, then attacks real samples: side channel measurements during cryptographic operations, voltage, clock and electromagnetic glitching, laser fault injection on the front and back of the die, invasive probing, and logical attacks on the software. Successful attacks are rated against the scheme's scale of attack potential, which reflects how much effort, expertise and equipment they needed. A product passes when the attacks that work need more than the target level allows.

What certification does and does not tell you

For a buyer, a certificate means a qualified lab tried hard to break the chip and failed within a defined budget, against a defined target. It does not cover attacks invented after the evaluation, firmware loaded later, or application code the chip vendor never saw. The Tangem case sits exactly in that gap: certified hardware, application firmware with a single check, and no way to update it. The practical rule for device makers is to treat a certified secure element as a strong foundation, not as a guarantee for the code you put on it.

How can an LED leak data or sound?

An LED wired to a device's power supply gets slightly brighter and dimmer as the device works harder or less hard. A photodiode or camera can record that flicker and recover what the device was doing: data on a modem line, speech from a speaker, even key material from a card reader. Filtering the supply stops it.

How it works

Every electronic device draws current that varies with its workload. A power or status LED connected directly to that supply follows the variation. The changes are far too fast and small for eyes, but a photodiode, a telescope with an electro-optical sensor, or a modern phone camera can capture them. Signal processing then turns brightness back into data, sound or a cryptographic trace. This is called an optical side channel.

Five-stage diagram from a device at work through changing power draw, an LED on the supply, a camera or sensor, to analysis, with a fix below each stage and three documented examples

Figure 7. When an LED tells on its device, and the fixes at each stage.

What the research showed

2002: status lights carrying data. Joe Loughry and David Umphress showed in Information Leakage from Optical Emanations (ACM Transactions on Information and System Security) that LED status indicators on data communication equipment could carry "a modulated optical signal that is significantly correlated with information being processed by the device." Many kinds of device were vulnerable, including modems and IP routers, and they described design changes that block the attack.

2021: speech from power LEDs. Ben Nassi and colleagues at Ben-Gurion University described the Glowworm attack, which recovers sound from the power indicator LED of devices connected to speakers. They recovered speech with good intelligibility from 15 metres and fair intelligibility from 35 metres, testing speakers, USB hub splitters and microcontrollers. CVE-2021-38548 records the issue for JBL Go 2 speakers: "The power indicator LED of the speakers is connected directly to the power line," so its intensity follows the power use, which follows the sound being played.

2023: keys from video. The same group's Video-Based Cryptanalysis, posted in 2023, used ordinary cameras. By exploiting a camera's rolling shutter, the way it reads the image line by line, they raised the effective sampling rate from 60 frames per second to about 60,000 measurements per second on an iPhone 13 Pro Max. They recovered a 256-bit ECDSA key from a smart card using video of its reader's power LED taken by an Internet-connected security camera 16 metres away, and a 378-bit SIKE key from a Samsung Galaxy S8 by filming the power LED of USB speakers on the same hub that charged the phone.

The impact

The leak only matters if the device handles something worth stealing and its cryptography leaks through power use in the first place. The 2023 attacks worked because the targeted cryptographic code drew power in a way that varied with secret values, a known class of weakness in some implementations. Speech recovery needs the LED to sit on the speaker's power path. In both cases the attacker also needs a clear, steady view of the LED, which is why a compromised camera already pointed at the device was part of the 2023 demonstration.

How to detect and defend

Detection is close to impossible from the victim's side, because the attacker only watches. Defence therefore belongs in design and in placement.

  • Filter the LED's supply. A capacitor, regulator or constant-current driver between the power rail and the LED removes the variation.
  • Drive indicator LEDs from firmware, not from the rail. An LED that is simply on or off says nothing about the workload.
  • Use cryptographic libraries hardened against power side channels, so the power trace does not depend on secret values.
  • Cover or disable LEDs on sensitive equipment where the design cannot change.
  • Secure the cameras that can see sensitive devices. A compromised security camera is a ready-made sensor. Cameras are networked computers with firmware that needs patching like any other.

Can infrared light blind or fool a security camera?

Yes, within limits. Bright visible or infrared light can saturate a camera's sensor and push its automatic exposure around, which hides what it should see, and infrared light can mislead some face recognition models. Modern cameras can raise tamper alerts, and overlapping cameras limit the damage.

How cameras handle light

Most security cameras have a day mode, with an infrared-cut filter in front of the sensor so colours look natural, and a night mode, where the filter moves aside and built-in infrared LEDs light the scene. Image sensors are sensitive to near-infrared light that people cannot see. Automatic exposure adjusts to the brightest parts of the scene, and analytics or recognition software runs on the processed image.

Three columns explaining how cameras see in day and night modes, how bright and infrared light disrupts them, and how operators detect and defend with image health alerts, overlapping views, placement and monitoring

Figure 8. Blinding and misleading cameras with light, and how operators respond.

What the research showed

Petit and colleagues' 2015 work tested an automotive camera, a MobilEye C2-270, against several light sources. They found a 5 mW 650 nm laser "the most effective way to temporary blind a camera," and in some cases "the camera never recovers from it." Infrared sources, an 850 nm LED spot and a 940 nm LED matrix, gave blinding times of more than five seconds. Their countermeasures apply to any camera: redundancy, because a narrow beam hits one camera at a time; a removable infrared-cut filter, which they note is already used on security cameras and blocks infrared light when applied; and photochromic lens materials that darken under intense light.

In 2018, Zhe Zhou and colleagues showed in Invisible Mask that infrared LEDs mounted on a cap could shine patterns on a face that people cannot see but a camera can. Against the FaceNet recognition system, they found adversarial examples implementable with infrared light more than 70 percent of the time, enough to evade recognition or impersonate someone from a photo of them. The lesson is less about cameras than about recognition models: a model trained only on visible-light faces can be steered by light it was never taught to expect.

The impact

Glare or infrared flooding usually causes a gap in coverage rather than a lasting compromise. That gap can still matter if it hides a person at the moment they pass. Face recognition errors matter more where recognition grants access, rather than simply tagging footage for review.

How to detect it

Many professional cameras can raise alerts when their view changes in ways that suggest tampering. Axis's Image Health Analytics, for example, detects blocked, redirected, blurred and underexposed images. Its manual is candid about the limits: it "can detect when something blocks the camera view, but it can't determine the intention," sudden lighting changes can trigger events, and a strong headlight can make a camera switch from night to day mode. Those limits are a reason to route alerts to a person who can look, not a reason to switch them off. Gaps in footage, sudden exposure changes and frequent day-night switching at one camera are all worth investigating.

How to defend against it

  • Overlap fields of view so no single camera is decisive.
  • Mount cameras high, use hoods, and avoid aiming them at streets, headlights or windows where bright light is common.
  • Turn on tamper or image health alerts and send them to the monitoring platform.
  • Do not let face recognition alone open doors. Pair it with a badge, PIN or liveness check, and test recognition systems against infrared light.
  • Patch camera firmware and secure camera accounts. A camera taken over through its software is worse than one dazzled for a few seconds, as the 2023 LED research showed.

How secure are laser links between buildings and satellites?

Free space optical links carry data on narrow infrared beams between buildings and between satellites. The narrow beam makes them much harder to intercept than radio, and weather or deliberate light usually causes an outage rather than altered data. The practical risks are the terminals at each end, so encrypt the traffic and secure the equipment.

Where laser links are used

Free space optical communication (FSO) sends data through air or space on light instead of fibre. Wikipedia's overview lists the classic use as LAN-to-LAN links between buildings on a campus or across a city, including across a road the operator does not own, and notes FSO's "license-free long-range operation (in contrast with radio communication)." Systems at 1550 nm can transmit more power than 850 nm systems while staying eye-safe.

Commercial deployments have grown. Taara, which began at X and spun out as its own company in 2025, says its Lightbridge reaches up to 20 Gbps over up to 20 km, that it has deployed hundreds of links in more than a dozen countries, and that it uses mirrors, sensors, precision optics and software to align the beam. In space, SpaceX's Starlink satellites pass traffic between each other on lasers. At SPIE Photonics West in January 2024, a SpaceX engineer described, as Hackaday reported, over 9,000 lasers in the constellation moving more than 42 petabytes a day, with 100 Gbps transceivers and links of up to 5,400 km.

Diagram of a building-to-building laser link and satellite laser links, with interception and interference considerations and four controls: encrypt the link, treat it as a conduit, secure the terminals, and monitor and fail over

Figure 9. Laser links between buildings and between satellites, and the controls that matter.

Eavesdropping considerations

A laser link's beam is narrow. As Wikipedia puts it, "The light beam can be very narrow, which makes FSO hard to intercept, improving security." To read the signal, an eavesdropper generally needs a receiver in the beam or very close to it, typically near the receiving end where the beam has spread most, and doing that without disturbing the link is difficult. Radio is different. A study by researchers at UC San Diego and the University of Maryland, described by Photonics Online, captured unencrypted geostationary satellite traffic, including corporate emails and utility telemetry, from a rooftop with about $800 of off-the-shelf equipment, because satellite radio beams cover wide areas. The editorial's point is that narrow optical beams make interception far less likely.

Far less likely is not impossible, and physics should not be the only control. Treat any link that leaves your building, optical or radio, as untrusted, and encrypt the traffic on it.

Interference considerations

Fog is the main enemy of terrestrial FSO; Wikipedia notes that fog "often prevents FSO laser links over 500 meters ... from achieving a year-round availability" for commercial service. Rain, snow, smog, building sway, misalignment and "interference from background light sources (including the sun)" also degrade links. Deliberate interference with bright light falls into the same category: it can cause an outage, but with encryption and integrity checks on the traffic it cannot silently change the data. That makes availability the planning problem, solved with monitoring and a backup path.

The terminals are the real attack surface

An FSO terminal is a networked computer with firmware, a management interface and often a cloud dashboard. A compromised terminal does not need anyone to intercept a beam: it sees all the traffic in the clear if the traffic is not encrypted end to end. That is the same exposure as any router or radio bridge, and the same controls apply.

Guidance for operators

For businesses and operators of industrial and commercial sites that use laser links between buildings, treat each link the way industrial security standards treat any connection between network areas.

  • Know what you have. Keep an asset inventory that includes each optical terminal, its firmware version and its management access.
  • Segment. In the Purdue model, industrial networks are organised in levels, from the physical process and field devices (levels 0 and 1) through control systems (level 2) and site operations (level 3) to business systems (level 4). IEC 62443 groups assets with common security requirements into zones and groups the communication paths between zones into conduits, each with its own protections. A laser link between buildings is a conduit: put a firewall at each end and allow only the traffic that zone pair needs.
  • Encrypt. Run MACsec, IPsec or TLS over the link so neither a tap nor a compromised terminal sees clear data.
  • Secure remote access. Manage terminals through a jump host or VPN with multi-factor sign-in, never from the open internet.
  • Monitor. Alert on link drops, signal strength changes, alignment events and configuration changes.
  • Plan for outages. Keep a fibre or radio backup path, backups of terminal configuration and an incident plan that covers losing the link.

Our explainers on IEC 62443 roles, zones and conduits and the Purdue model and where it falls short go deeper.

Why are laser pointers a safety problem for aviation?

A handheld laser aimed at an aircraft can cause glare, flash blindness and after-images in the cockpit during the busiest phases of flight. Pilots' associations describe it as a serious safety issue. Crews are trained to look away, hand control to the other pilot and report it.

This one is not a device hack, but it is the most common real-world misuse of lasers against transport, and it belongs in any honest account of laser risk. This section draws only on pilots' own organisations.

The International Federation of Air Line Pilots' Associations (IFALPA) says in its 2021 position paper that thousands of laser strikes are reported by pilots every year in the United States alone, and that strikes "can result in distraction, disruption, disorientation and, in extreme cases, blindness and incapacitation." Most happen near the ground, but events have been recorded above 8,000 feet. IFALPA notes that "so far, when a pilot has been attacked on final approach, they have been able to hand control to the other pilot to complete the landing safely," and warns that more powerful lasers could affect both pilots.

IFALPA's medical briefing explains the effects. Glare disrupts vision while the light is present. Flash blindness is a temporary loss of vision that persists after the light is gone. After-images linger in the visual field. It states that a common 5 mW laser pointer "can easily cause glare and distract pilots up to 3700ft," and that retinal damage from laser incidents in aviation is rare. Its recommended crew actions are to look away and shield the eyes, avoid rubbing them, check whether the other pilot was exposed and hand over control if not, manoeuvre away from the beam where possible, and report the event.

IFALPA's position is that the best defence is preventing attacks on the ground, through effective rules, enforcement and public awareness campaigns. For everyone else, the advice is short: never point a laser at an aircraft, a vehicle or a person, and report it if you see someone doing it.

How should you weigh these risks?

Weigh them by what the attacker needs and what they gain. Laser pointers at aircraft are common and dangerous. Voice injection and camera glare are cheap but limited by placement and confirmation. LiDAR spoofing has high potential impact but was shown in labs. Laser fault injection is powerful but costly and one device at a time.

Risk versus effort chart placing laser pointers at aircraft, Light Commands, camera glare, LED side channels, LiDAR spoofing, laser fault injection and free space link interception by attacker effort and potential impact

Figure 10. How much effort, how much harm. Positions are qualitative judgements from the published work.

A few principles help.

  • Line of sight is a strong control. Most of these attacks need a clear path to a sensor. Placement, covers and shrouds cost little and remove a lot of risk.
  • Impact depends on what the sensor controls. A voice assistant linked to nothing important is a nuisance. One linked to a front door without a PIN is a security problem. Fix the link, not the light.
  • Cost and access cut both ways. A $250,000 lab and the device in hand put laser fault injection out of reach for most attackers, but well within reach of someone targeting a stolen high-value device. Match the countermeasures to what the device protects.
  • Lab results are early warnings. LiDAR and camera research shows where perception systems need plausibility checks and fusion before they carry more responsibility.
  • Software decides the outcome. In nearly every case, the step that turns a signal into harm is firmware or application code. That is the cheapest place to add a check.

How do you detect light-based attacks?

Light attacks are hard to see directly, so detection relies on what the device does: unexpected voice commands, sensor disagreements, camera image health alerts, fault counters in secure chips, and optical link drops. Log these centrally and look at patterns by place and time.

Signals for home users

  • Voice assistant history showing commands nobody gave, or lights and replies when nobody spoke.
  • A small bright spot on a smart speaker or camera, or a laser dot through a window.
  • Smart lock, garage door or purchase notifications you did not trigger.
  • Camera footage with sudden white-outs, unexplained day-night switching or gaps.
  • A hardware wallet or key card that is missing or shows signs of having been opened.

Signals for businesses and operators

  • Camera image health and tamper events, especially several at one location.
  • Voice device logs in meeting rooms and lobbies, if voice assistants are allowed at all.
  • Optical link alarms: drops, falling signal strength, alignment changes or configuration edits outside change windows.
  • Card reader and payment terminal tamper alerts, and readers returned with damaged casings.
  • For vehicle fleets, logs of sensor disagreements, safety stops and remote interventions, mapped by location.

Signals for manufacturers

  • Field telemetry from multi-microphone fusion that rejects single-microphone commands.
  • LiDAR returns arriving outside expected timing windows, or replies to skipped pulses.
  • Secure element fault counters and lock events reported through device management.
  • Spikes in returned or failed devices with opened packages, which can mean someone is characterising your chip.
  • Vulnerability reports from researchers. Almost every case in this guide reached the vendor that way.

Why is this a firmware and software supply chain problem?

Every defence in this guide, from voice confirmation and microphone fusion to LiDAR plausibility checks, camera tamper analytics and fault-hardened secure element code, is software. That software is built from chip vendor SDKs, open source libraries and supplier firmware, and it has to be tracked, tested and updated like any other code.

Consider what sits between the light and the harm in each case.

  • A smart speaker runs a Linux-based or RTOS firmware image, audio processing from a chip vendor, a voice assistant client and a smart home integration layer. Microphone fusion and command confirmation live there.
  • A LiDAR has its own firmware that controls pulse timing. The perception stack that consumes its data may be built on open source frameworks such as Apollo or Autoware, on machine learning libraries, and on the carmaker's or supplier's own code. Randomised timing, plausibility checks and fusion live there.
  • A security camera runs embedded Linux, an image pipeline from the chip vendor, analytics, a web interface and a cloud agent. Tamper detection and the camera's own security live there.
  • A secure element runs a small operating system and applet code, often from a different company than the silicon maker, on top of the vendor's certified crypto library. The double checks the Donjon team recommended live there, and so does the update mechanism the Tangem cards lacked.
  • An optical link terminal runs embedded firmware and a management stack, often with a cloud dashboard.

Each of these is assembled from parts made by other companies. The firmware images contain open source components with their own vulnerabilities. Chip SDKs are copied into many products, as our post on how computer chips are made, from sand to software explains. And the companies who ship the final device often receive binaries, not source. That is the same supply chain problem our guides on laptops, cars and smart home devices describe, applied to the code that decides what a sensor's signal means.

Two lessons stand out. First, updatability is a security control. A LiDAR firmware that can be updated can gain randomised timing; a voice assistant can gain confirmation prompts; a card without an update path keeps its weakness for life. Second, you cannot fix what you cannot see. A maker who does not know which audio library, which perception framework or which crypto library version is in each product cannot answer a researcher's report quickly, or know which other products share the weakness. A software bill of materials (SBOM) for every firmware image and app is the starting point, as our guide to SBOMs for firmware and IoT devices explains.

How should home users protect themselves?

Keep smart speakers out of view of windows, require a PIN or phone confirmation for anything that unlocks, opens, pays or starts, check voice and camera history, install firmware updates, and keep hardware wallets and key cards in your possession.

  • Place voice assistants away from windows and out of sight lines from outside. Close blinds in rooms where they sit.
  • Turn on confirmation for sensitive actions. Use a voice PIN, voice match or phone approval for locks, garage doors, purchases and car functions, or do not link them to voice at all.
  • Review voice assistant history now and then, and turn on notifications for smart lock and garage door events.
  • Turn on camera tamper or image health alerts if your camera offers them, and mount outdoor cameras so they do not face streets or headlights.
  • Install firmware updates for smart speakers, cameras, routers and hardware wallets. Updates are how vendors add the software defences this guide describes.
  • Keep hardware wallets, payment cards and key cards physically safe, and treat a lost one as compromised: move funds, cancel cards, revoke keys.
  • Cover LEDs on sensitive equipment if you are concerned about optical leaks; a piece of opaque tape works.
  • Never point a laser at an aircraft, a vehicle or a person. It is dangerous and it is an offence in many countries.

How should businesses and operators protect their systems?

Inventory every sensor, camera, voice device, card reader and optical link; segment networks into zones and conduits; secure remote access to device management; monitor tamper and link alerts; keep backups and an incident plan; and buy certified secure hardware for keys and payments.

  • Asset inventory. List voice devices, cameras, card readers and payment terminals, LiDAR-equipped vehicles or robots, and optical link terminals, with firmware versions and owners. You cannot patch or monitor what is not on the list.
  • Segmentation. Use the Purdue levels and IEC 62443 zones and conduits to separate cameras, building systems, operational technology and business networks. Treat optical links and wireless bridges as conduits with firewalls at both ends.
  • Secure remote access. Manage cameras, readers and link terminals through a controlled path with multi-factor sign-in and logging. Close direct internet exposure.
  • Monitoring. Send camera image health events, reader tamper alerts, link alarms and fleet sensor disagreements to your monitoring platform, and assign someone to look.
  • Backups and resilience. Keep backup network paths for optical links, configuration backups for devices, and overlapping camera coverage in sensitive areas.
  • Incident response. Include physical tampering in your incident plan: who inspects a damaged reader, who reviews a camera gap, who replaces a lost key card.
  • Procurement. Buy secure elements, payment terminals and smart cards with recognised certifications such as Common Criteria, EMVCo or PSA Certified, and ask device vendors for SBOMs and their update policy.
  • Policy for voice devices. Decide whether voice assistants belong in offices, lobbies and meeting rooms at all. If they do, keep them away from windows and unlinked from anything that opens or pays.

How should manufacturers build devices that resist light-based attacks?

Threat model light for every sensor and chip, confirm sensitive actions in software, fuse sensors and check plausibility, harden firmware against faults, filter LED supplies, certify secure hardware with independent labs, and ship every device with an SBOM and a signed update path.

Three-column hardening checklist for home users, businesses and operators, and manufacturers

Figure 11. Hardening checklists for the people who use, operate and build these devices.

Sensors and perception

  • Threat model light explicitly for microphones, LiDARs, cameras and photodiodes, not only radio and network inputs.
  • Confirm sensitive voice actions and fuse microphone channels so a single-microphone command is rejected.
  • Randomise or skip LiDAR pulses, probe more than once where scan rate allows, and add physical consistency checks such as occlusion reasoning.
  • Fuse sensors and define safe fallbacks when they disagree.
  • Test against light. Include laser and infrared interference in sensor validation and in recognition model testing.

Chips and firmware

  • Choose certified secure elements for keys and payments, and read the certificate's scope.
  • Harden the code you put on them. Check every security decision more than once, encode flags with large Hamming distances, verify results before release, and handle fault detections safely.
  • Keep a signed update path for firmware, including applet code on secure elements where the platform allows it.
  • Use cryptographic libraries hardened against side channels and keep them current.

Hardware design

  • Filter LED supplies or drive indicator LEDs from firmware, not from the power rail.
  • Shield microphone ports with opaque materials and design enclosures that block direct light paths to sensors.
  • Add tamper evidence to devices that hold secrets.

Supply chain and disclosure

  • Build and keep an SBOM for every firmware image, app and backend, and require SBOMs from chip and module suppliers.
  • Watch for new vulnerabilities in what you have shipped and rank them by whether the vulnerable code is reachable.
  • Run a vulnerability disclosure programme and respond to researchers. Every case in this guide was published by researchers who could have been talking to you first.

How Safeguard helps teams that build these devices

Safeguard covers the software side of these devices in one platform: SBOMs for firmware and apps, binary and image scanning, vulnerabilities ranked by reachability, SCA, SAST and secrets scanning on code you own, fix pull requests gated by your CI, Gold patched packages and compliance evidence, as SaaS, in your VPC or on premises.

Safeguard does not change how a microphone diaphragm or a transistor reacts to a laser. It covers the part of the problem this guide keeps returning to: the firmware and software that decide what a device does with the signal, and the third-party components inside them. Device teams often stitch that together from a source scanner, a binary analysis tool, an SBOM manager and a GRC platform. Safeguard does it end to end, with one inventory and one vulnerability model from the first SBOM to the merged fix and the audit record.

Flow diagram of Safeguard for device makers: source repositories, binaries and images, extracted firmware and supplier SBOMs feed SBOMs, reachability-ranked vulnerabilities and code scanning, leading to fix pull requests, Gold patched packages, compliance evidence and SBOM delivery

Figure 12. Where Safeguard fits for device makers.

See inside the binaries and images you ship and receive

In the Safeguard console, Integrations > Upload App Binary accepts EXE, MSI and DLL files, Android APKs and iOS IPAs, QCOW2, VMDK, VHD and OVA images, and container tarballs. Safeguard identifies the embedded libraries and their versions, maps them to known vulnerabilities and records licences, and the result becomes a project with a full SBOM. For the devices in this guide, that covers the companion apps for smart speakers, cameras and hardware wallets, Linux-based images for cameras, gateways and link terminals, and the containers behind device clouds. Raw firmware images for microcontrollers and secure elements may need to be extracted into their component files first, and heavily packed or encrypted images can give incomplete results. Our guide to scanning app binaries with Safeguard and our explainer on binary composition analysis go further.

Rank vulnerabilities by what can actually be reached

A camera or LiDAR firmware image can carry hundreds of components. Safeguard's reachability analysis ranks the vulnerabilities your code can actually reach above those it cannot, with EPSS and known exploited context, so a team can tell the image parser that needs fixing this week from an unused function in a bundled library.

Scan the code you own

For firmware, sensor drivers, perception code, voice integration and device cloud services that you write, Safeguard runs software composition analysis (SCA) across the full dependency tree, static application security testing (SAST) on your own code, and secrets scanning that catches signing keys and credentials before they ship. A leaked firmware signing key undermines the update path that every defence in this guide depends on.

Fix it without waiting for upstream

For code in your repositories, Griffin, Safeguard's remediation model, opens a fix pull request that your own CI builds and tests before anyone merges it, under your policies. For open source components where upstream has not shipped a fix, or the fix needs a version you cannot move to yet, Safeguard Gold provides patched, hardened builds of those packages, and Safeguard can fork and patch the vulnerable package so your next firmware build pulls a fixed version. Closed source chip SDKs still have to come from their vendor, but the open source around them no longer has to wait.

Compliance evidence and deployment

The same SBOMs, vulnerability records and fix history feed Safeguard's compliance suite for frameworks such as SOC 2 and ISO 27001, and give device makers the raw material for customer reviews and for obligations such as the EU Cyber Resilience Act's SBOM requirements. Safeguard runs as SaaS, in your own VPC or fully on premises, in Zero AI mode or with a model you choose.

How this compares

Device teams commonly combine a source SCA tool such as Snyk or Black Duck, a separate binary analysis tool for firmware, and a GRC platform for audits. Those are capable products. Safeguard's case is consolidation and remediation: one inventory across source, binaries and images, one reachability-aware vulnerability model, and fix pull requests and patched packages rather than tickets. It can also run alongside an existing tool and take over SBOM delivery and fixing. For hardware-level resistance to lasers, you still need certified silicon and an evaluation lab; Safeguard covers the software you put on top.

The takeaway

Light-based attacks sound exotic, and some of the equipment is. But the research tells a consistent, calm story. A microphone that hears a laser, a LiDAR that records an echo it never sent, a camera that adjusts to a beam, a chip that skips one check under a pulse of light: in each case the sensor or circuit behaved as physics says it will, and the harm came from code that trusted it.

That is good news for defenders, because code is where the cheapest fixes are. Ask before unlocking. Compare microphones. Check that an obstacle casts the shadow it should. Verify a security decision twice. Filter an LED. Keep a device updatable, and know what is inside every firmware image you ship. Labs will keep finding new ways for light to reach silicon. The teams that handle them well will be the ones who can see their software supply chain and fix it quickly.

Try it yourself

Click through these interactive Safeguard demos to see the steps described in this guide.

Test yourself

Why do sensors and chips respond to light at all?

Silicon reacts to light. Enough light on a transistor or sensor chip creates a small current, and sensors are built to turn tiny changes into signals, so they turn unwanted light into signals too.

What did Light Commands show?

A laser whose brightness follows a voice recording, aimed at a smart speaker's microphone port, makes the microphone report that voice. Google, Amazon, Apple and Facebook voice devices accepted the commands, in one case from 110 metres and through a window.

How can LiDAR be spoofed?

Light from someone else, timed to arrive like an echo, can add points for an obstacle that is not there or drown out real echoes so an obstacle disappears. Researchers showed both in labs and simulators, and sensor fusion, randomised timing and plausibility checks push back.

What is laser fault injection?

Precisely timed laser pulses at an opened chip make transistors briefly conduct, flipping a bit or skipping an instruction. It needs the chip in hand, specialist equipment and expertise, and chips resist it with shields, light sensors, redundancy and hardened firmware.

How can an LED leak information?

An LED wired to a device's power supply gets slightly brighter and dimmer as the device works harder or less hard. A photodiode or camera can record that flicker and recover what the device was doing. Filtering the supply stops it.

What simple steps help home users defend against light-based attacks?

Keep smart speakers out of view of windows, require a PIN or phone confirmation for anything that unlocks, opens, pays or starts, check voice and camera history, and install firmware updates.

Frequently asked questions

Can a laser hack Alexa, Google Home or Siri?

Researchers showed in 2019 that a laser whose brightness follows a voice recording, aimed at a device's microphone port, can make Alexa, Google Assistant, Siri and Portal devices accept commands, from up to 110 metres and through a window. It needs line of sight and steady aim. Keep speakers away from windows, require a PIN or phone confirmation for locks, purchases and cars, and review the assistant's history.

Can a laser make a self-driving car see things that are not there?

In labs and simulators, yes. Researchers have used light timed against a LiDAR's pulses to add phantom obstacles and, since 2023, to remove real ones from the point cloud. These attacks need precise timing and line of sight and were demonstrated in controlled settings. Vehicles defend with sensor fusion, plausibility checks, randomised LiDAR timing and safe fallbacks when sensors disagree.

What is laser fault injection?

Laser fault injection fires precisely timed laser pulses at an opened chip so transistors briefly conduct, flipping a bit or skipping an instruction such as a password check. Evaluation labs use it to test smart cards and secure elements before certification. It needs the chip in hand, specialist equipment that can cost hundreds of thousands of dollars, and expertise. Chips resist it with shields, light sensors, redundancy and hardened firmware.

Can the LED on a device leak data?

It can when the LED is wired directly to the power supply, because its brightness then follows the device's power use. Research has recovered network data from modem LEDs, speech from speaker power LEDs at up to 35 metres, and a smart card key from video of a reader's power LED. Makers fix it by filtering the LED's supply; users can cover LEDs on sensitive equipment.

Can infrared light blind a security camera?

Bright visible or infrared light can saturate a camera's sensor and confuse its automatic exposure, usually for seconds, and infrared patterns can mislead some face recognition models. Use cameras with tamper or image health alerts, overlap fields of view, mount cameras away from streets and headlights, and do not let face recognition alone open doors.

Are laser links between buildings and satellites secure?

They are much harder to intercept than radio because the beam is very narrow, and interference usually causes an outage rather than altered data. The practical risk is the terminal at each end, which is a networked computer. Encrypt traffic over the link, treat it as a conduit with firewalls at both ends, patch and lock down terminals, and keep a backup path.

Why is shining a laser pointer at a plane dangerous?

A laser in the cockpit can cause glare, flash blindness and after-images during take-off and landing. IFALPA, the international pilots' federation, says even a 5 mW pointer can cause glare up to 3,700 feet, and that strikes can cause distraction, disorientation and in extreme cases incapacitation. Pilots are trained to look away and hand over control. Never aim a laser at an aircraft.

How do chip makers prove their chips resist laser attacks?

They send chips to accredited labs that attack them, including with laser fault injection and side channel analysis, under schemes such as Common Criteria (ISO/IEC 15408), EMVCo for payment chips, SESIP (EN 17927) and PSA Certified. A certificate shows the chip resisted a defined level of attack. It does not cover firmware added later, which is why that code needs its own hardening and an update path.

Sources

Never miss an update

Weekly insights on software supply chain security, delivered to your inbox.

We email you a link to confirm, and only then add you to the list. Unsubscribe in one click. Privacy

Self-healing security runs on Safeguard.

Your first fix PR is minutes away.

No sales call required, even your agent can complete the purchase over MCP.