Safeguard
Supply Chain Security•

How Life Sciences Companies Get Hacked

Life sciences cybersecurity: how pharma plants, biotech labs, sequencers, clinical trials and genomic data get hacked, with real cases and fixes.

Safeguard Security Team
60 min read

Life sciences companies are rarely hacked through anything exotic. The documented cases are ransomware that spreads from office IT into plants and trial systems, lab instruments running old Windows and exposed vendor services, customer accounts opened with reused passwords, and open source packages or AI models that run code the moment a pipeline loads them. Each one has a known defence, and this guide walks through them in order.

Pharma manufacturers, biotech labs, contract research organisations, genomics firms and the makers of lab instruments all run on software that most of their staff never see. A sequencer is a Windows PC with a camera and chemistry attached. A vaccine plant is a stack of controllers, operator screens and a manufacturing execution system, tied to the business network for orders and batch release. A clinical trial lives in a vendor's cloud. A drug discovery team pulls hundreds of Python packages and pretrained models from the internet every week.

This guide is for four groups. IT and security teams at pharma and biotech companies. Quality, validation and automation engineers who own GMP systems and have to keep them both secure and compliant. Research computing and bioinformatics teams. And product security teams at lab instrument and lab software makers, who will find the supply chain, SBOM and hardening sections most useful. Jargon is defined on first use.

This is a defender's guide written from the attacker's point of view. For each attack class you will find how it works and why, what an attacker needs, documented cases with CVE numbers where they exist, the impact, how to detect it and how to defend against it. There are no exploit steps, payloads or tool command lines, and no attribution of incidents to any country or group. A note on the law: testing systems or instruments you do not own, or do not have written permission to test, is illegal in most countries. In a plant or a lab it can also spoil a batch or a run. Tools and research teams are named only as context for authorised work.

This post is part of a device and industry security series. The companion guides cover how Wi-Fi gets hacked, how laptops get hacked and how medical devices and wearables get hacked. For where the boards and firmware inside instruments come from, see how a laptop is built and fabless chip designers and the software supply chain.

Contents

  1. How do life sciences companies get hacked? The short answer
  2. What software runs a life sciences company?
  3. What is the attack surface of a life sciences company?
  4. Attack 1: Can lab instruments and DNA sequencers be hacked?
  5. Attack 2: How are LIMS, ELN and lab data systems attacked?
  6. Attack 3: How does ransomware stop pharma manufacturing?
  7. Why are validated GMP systems so hard to patch?
  8. Attack 4: How are clinical trial platforms attacked?
  9. Attack 5: How does genetic data leak?
  10. Attack 6: How is research and manufacturing IP stolen?
  11. Attack 7: How do open source packages and Python dependencies compromise research pipelines?
  12. Attack 8: What are the security risks of AI models in drug discovery?
  13. Which attacks matter most for life sciences?
  14. How can you detect an attack in a lab, plant or research pipeline?
  15. How do the defences fit together?
  16. Hardening checklists for biotech labs, pharma manufacturers and instrument makers
  17. Where does the life sciences software supply chain break?
  18. How Safeguard helps life sciences companies and instrument makers
  19. The takeaway
  20. Frequently asked questions
  21. Sources

How do life sciences companies get hacked? The short answer

Most life sciences breaches start in ordinary IT: a phished login, an exposed remote access service or a reused password. From there, attackers reach lab instruments, plant systems and trial platforms that cannot be patched quickly because they are validated. Software supply chain attacks through updates, open source packages and AI models are the fastest growing route.

The documented record falls into eight attack classes, each covered below:

  1. Lab instruments and sequencers. Instrument PCs run vendor services with network-facing flaws, old operating systems and, in at least one documented case, firmware without Secure Boot.
  2. LIMS, ELN and lab data systems. Web applications that hold every sample and result, with ordinary web vulnerabilities and high value for ransomware.
  3. Ransomware in manufacturing. Malware that crosses from office IT into plant operations and stops production.
  4. Clinical trial platforms. Shared cloud services where one vendor's outage stops many sponsors' trials at once.
  5. Genetic data. Consumer and research accounts opened with reused passwords, then amplified by sharing features.
  6. Intellectual property theft. Quiet, long-running access to research data, protocols and process know-how.
  7. Open source in research pipelines. Malicious packages and compromised build pipelines in the Python and R ecosystems that bioinformatics depends on.
  8. AI models in drug discovery. Model files that execute code when loaded, plus data leakage and poisoning risks.

Two facts make life sciences different from most industries. First, many systems are validated: documented and tested to prove they do what they should, under quality rules that require any change to be controlled and tested again. That makes patching slow. Second, the data has an unusually long life. A stolen password can be changed. A stolen genome, a trial dataset or a manufacturing process cannot.

What software runs a life sciences company?

A life sciences company runs four stacks: lab instruments with embedded Windows PCs, lab informatics such as LIMS and ELN, manufacturing operational technology (OT) under GMP, and research and clinical platforms in the cloud. Most layers in each stack come from suppliers, so most vulnerabilities arrive through software the company did not write.

Four-column diagram of the software stacks for lab instruments, lab informatics, manufacturing OT, and research and clinical platforms, coloured by whether each layer is the company's own code, supplier code, operating system or hosted SaaS

Figure 1. Four typical stacks. In each one, most layers come from suppliers.

Lab instruments

Sequencers, mass spectrometers, liquid handling robots, plate readers and clinical analysers are computers with specialised hardware attached. Inside most of them is an industrial PC board, often made by an outside board supplier, running Windows or Linux. On top of that sit the instrument vendor's control and analysis software and a handful of vendor services: run managers that expose a small web interface, file copy services that move run data to network storage, and remote support agents that let the vendor's field engineers diagnose problems.

The instrument is usually bought, installed and qualified as one unit. The lab's IT team often cannot patch the operating system without the vendor's approval, because an untested update could change how the instrument behaves. That is the root of most instrument security problems covered below.

Lab informatics: LIMS, ELN and CDS

A LIMS (laboratory information management system) tracks every sample: where it came from, which tests ran, who ran them and what the results were. An ELN (electronic lab notebook) replaces paper notebooks for experiments. A CDS (chromatography data system) controls chromatography instruments and stores their raw data. These are typically web applications on Java, .NET or Python stacks, backed by SQL databases and file shares. They are the system of record for regulated data, so their integrity matters as much as their availability.

Manufacturing OT under GMP

GMP (good manufacturing practice) is the set of quality rules that drug and biologics manufacturing must follow. It belongs to a family of quality guidelines known collectively as GxP, where the x stands for manufacturing, laboratory, clinical and so on; their shared purpose is to ensure a product is safe and meets its intended use (Wikipedia, GxP).

A pharma or biologics plant runs on layers of OT:

  • Sensors and actuators measure temperature, pressure and pH, and drive valves, pumps and scales.
  • PLCs (programmable logic controllers) run the control logic for bioreactors, clean-in-place systems and filling lines.
  • SCADA and HMI systems give operators screens to watch and steer the process, and a historian records process data over time.
  • An MES (manufacturing execution system) runs batch production: recipes, electronic batch records, weigh and dispense, and material tracking. The ISA-95 standard places the MES at Level 3, between business planning systems at Level 4 and process control at Levels 0 to 2 (Wikipedia, MES).

Computerised systems that support GMP work are validated. The industry's leading guidance on how to do that is ISPE's GAMP 5, now in its second edition, subtitled A Risk-Based Approach to Compliant GxP Computerized Systems (ISPE). In the United States, 21 CFR Part 11 sets the criteria under which electronic records and electronic signatures are treated as trustworthy and equivalent to paper; in practice it means system validation, audit trails, access controls and electronic signature controls for the systems that hold regulated records (Wikipedia, Part 11). This guide uses these terms only as the industry uses them. It is not compliance advice.

Research and clinical platforms

Research computing runs bioinformatics pipelines: chains of open source tools, scripts and notebooks that turn raw instrument output into variant calls, expression profiles or protein structures. They are built from packages on PyPI, conda channels, Bioconductor and CRAN, packed into containers and run by workflow engines on cloud compute. Drug discovery teams increasingly add AI models, some trained in house and many downloaded from public model hubs.

Clinical trials run on hosted platforms: EDC (electronic data capture) for case report forms, eCOA (electronic clinical outcome assessment) for patient diaries and questionnaires, IRT (interactive response technology) for randomisation and drug supply, and safety databases. These are usually multi-tenant SaaS from a small number of specialist vendors.

What is the attack surface of a life sciences company?

The attack surface is every route by which outside input reaches these four stacks: staff and customer accounts, remote access, lab and plant networks, vendor support channels, software updates, open source packages, public AI models and cloud storage. Supplier routes now outnumber the routes a company controls directly.

Map of twelve entry points around a life sciences company, coloured by whether they depend on people and accounts, sit on the internal network, arrive through a supplier, or face the internet

Figure 2. Twelve entry points. Half of them arrive through a supplier.

Grouped by what an attacker needs to reach them:

  • People and accounts. Staff email, VPN and remote desktop logins, and customer or patient accounts on consumer genomics and patient portals. Reached with phishing, stolen or reused passwords, and tired help desks.
  • Internal networks. Lab instruments on the lab network, LIMS and ELN servers, and the plant network with its MES, SCADA and PLCs. Reached by anyone who already has a foothold in office IT, if the networks are not separated.
  • Suppliers. Clinical trial SaaS vendors, instrument vendors' remote support, trusted software update channels, open source packages and containers, and public AI models. Reached by compromising the supplier or publishing something malicious where the company will fetch it.
  • The internet. Cloud storage buckets, research pipeline endpoints and any service exposed by mistake.

The rest of this guide takes the documented attacks one at a time. Figure 3 puts them on a single timeline.

Timeline from June 2017 to June 2025 of documented life sciences security events, coloured by vulnerability research, ransomware and disruption, open source and AI supply chain, and genetic data privacy

Figure 3. Nine years of documented events. Ransomware and supply chain incidents dominate.

Attack 1: Can lab instruments and DNA sequencers be hacked?

Yes. Researchers and vendors have documented critical vulnerabilities in sequencer software services, including flaws rated CVSS 10 that allowed remote code execution, and sequencer firmware booting a 2018 BIOS without Secure Boot. No exploitation in the wild has been reported for these cases, and the vendor issued fixes.

Layered diagram of a networked sequencer showing the control software, vendor network services, Windows operating system and supplier motherboard firmware, with the documented 2022, 2023 and 2025 vulnerabilities at each layer

Figure 4. Four layers on one instrument PC. Documented flaws have been found at three of them.

How it works and why

An instrument PC sits on the lab network so it can receive sample sheets, send run data to storage and the LIMS, and let the vendor support it. Each of those jobs is done by a service listening on the network. If a service binds to every network interface instead of just the one it needs, accepts requests without authentication, or runs with administrator rights, then anyone who can reach the instrument over the network may be able to run code on it.

Below the software sits firmware. BIOS and its modern successor UEFI are the first code a PC runs when it powers on. Secure Boot checks that boot code is signed and unchanged. Firmware write protection stops software on the running system from rewriting the firmware chip. Without them, malware that gets administrator rights can install itself below the operating system, where reinstalling Windows will not remove it.

Why does this happen? Instruments are designed to run for ten years or more, their PCs are bought from board suppliers, and their software is validated as a whole. The incentive is to change as little as possible once an instrument ships.

What the attacker needs

Network access to the instrument, which in practice means a foothold somewhere on the same flat network: a phished workstation, a compromised file server or a misconfigured remote access path. For the firmware attack, administrator access on the instrument PC first. None of these flaws are reachable from the internet unless the instrument has been exposed by mistake.

Real cases

Illumina Local Run Manager, 2022. Local Run Manager (LRM) is the software Illumina sequencers use to set up and analyse runs. In 2022, five vulnerabilities in LRM were published. CVE-2022-1517, CVE-2022-1518 and CVE-2022-1519 carried CVSS scores of 10 and allowed remote code execution at operating system level (LRM ran with elevated privileges), uploads outside the intended directories, and upload of arbitrary files. The others allowed connection without authentication (CVE-2022-1521) and interception of unencrypted traffic (CVE-2022-1524). Affected instruments included the NextSeq 550Dx and MiSeq Dx diagnostic instruments and the NextSeq 500, NextSeq 550, MiSeq, iSeq 100 and MiniSeq research instruments. Illumina issued updates to prevent remote exploitation and told affected healthcare customers on 3 May 2022 (SecurityWeek).

Illumina Universal Copy Service, 2023. The Universal Copy Service (UCS) copies run data off the instrument. In April 2023 two vulnerabilities were disclosed. CVE-2023-1968, rated CVSS 10.0, allowed the service to bind to an unrestricted IP address and so listen on every interface that could accept remote connections. CVE-2023-1966, rated CVSS 7.4, came from the service running with unnecessary privileges, which could let an attacker execute code at operating system level and change settings, software or data. They affected iSeq, MiniSeq, MiSeq, NextSeq and NovaSeq systems (SecureWorld). The advisory warned that exploitation could affect the accuracy of genomic results, producing no results, false results or modified results, or exposing data. Mitigations included updating to the patched version, configuring UCS account credentials, keeping instruments off the internet and using secure remote access (JD Supra). No public exploits were known at the time.

iSeq 100 firmware, 2025. In January 2025, Eclypsium researchers published an analysis of the iSeq 100. The instrument booted a BIOS build from 2018 (B480AM12, dated 04/12/2018) with known vulnerabilities, ran in CSM (Compatibility Support Mode, which lets UEFI boot legacy BIOS code), had firmware read and write protections disabled and did not use Secure Boot. An attacker with access to the system could overwrite the firmware to brick the instrument or plant a persistent implant (Eclypsium). The analysis linked nine vulnerabilities to five issues, including exposure to LogoFAIL, a 2023 class of UEFI image parsing flaws (BleepingComputer). The motherboard came from IEI Integration Corp., which also builds boards for other medical and industrial equipment, so the researchers said similar issues were likely in other devices. Eclypsium saw no evidence of exploitation and confirmed that Illumina had provided a fix and notified customers. Illumina said its initial evaluation found the issues were not high risk and that it was working to improve how it delivers security updates to instruments in the field (The Register, The Record).

Old Windows on instruments. Illumina's own FAQ for moving instruments from Windows 7 to Windows 10 says instruments continue to function on Windows 7, but that using an unsupported operating system increases the risk of exposure to malware, and that customers who stay on Windows 7 should take steps to minimise exposure. Upgrades are arranged through Illumina technical support (Illumina Knowledge). The same pattern holds across the instrument market: the vendor, not the lab, controls the upgrade path.

These cases involve one vendor because that vendor's products have been studied publicly, and because Illumina published fixes and guidance. Treat them as representative of instrument PCs in general, not as a judgement on one company.

Impact

The serious impact is on data integrity rather than data theft. A sequencer or analyser that has been tampered with can produce results that look valid and are not. In research, that wastes months. In a clinical or quality control setting, it can affect a diagnosis or a batch release. The second impact is availability: an instrument bricked by a firmware overwrite needs a field engineer, and ransomware that encrypts an instrument PC stops the runs on it.

Detection

  • Keep a live inventory of every instrument PC: model, serial, operating system version, vendor software versions and which network services it exposes.
  • Watch the instrument network for new listening ports, new outbound destinations and logins at odd hours. Instruments are predictable, so anything new stands out.
  • Compare vendor service versions against the vendor's security advisories on a schedule. Most instrument vendors now publish them.
  • Where the vendor supports it, check firmware versions and Secure Boot status during periodic maintenance.

Defence

  • Segment instruments onto their own network, with only the connections each instrument needs: to its storage share, the LIMS, and a vendor remote support gateway. No direct internet.
  • Apply vendor security updates promptly through the vendor's process. Ask your vendor for its update schedule in writing.
  • Restrict and authenticate services. Configure service credentials where the vendor allows, as Illumina recommended for UCS.
  • Control remote support. Vendor access through a jump host with multi-factor authentication (MFA) and session recording, enabled only when needed.
  • Plan operating system upgrades with the vendor before end of support, and budget for them at purchase.
  • Buy with security in the contract. Ask for an SBOM, a supported operating system roadmap, Secure Boot enabled, and a vulnerability disclosure policy before you sign.

Attack 2: How are LIMS, ELN and lab data systems attacked?

LIMS, ELN and chromatography data systems are web applications attacked like any other: injection flaws, weak authentication, exposed admin interfaces and unpatched libraries. The bigger risk is ransomware on the servers and file shares behind them, which can halt sample processing for weeks, as a 2024 pathology lab attack showed.

How it works and why

A LIMS is a database with a web front end, plugins for instruments and integrations with ERP and quality systems. Many are heavily customised, and upgrades are expensive because each one needs revalidation. So older versions stay in service, and with them older libraries and older bugs. The data is highly structured and critical, which makes it a prime ransomware target: encrypt the LIMS and its file shares, and a lab cannot release results.

SQL injection is a common example. It happens when an application builds a database query from user input without separating the two, so specially crafted input changes the query itself.

What the attacker needs

For web flaws, network access to the application and often a valid low-privilege login. For ransomware, a foothold in the corporate network and enough privilege to reach the servers and their backups.

Real cases

LabVantage LIMS, CVE-2024-5771. Published in June 2024, this is an SQL injection vulnerability in a request handler of LabVantage LIMS 2017, reachable over the network by a user with a low-privilege account. It is rated medium, CVSS 3.1 base score 6.3, and its details were made public (Tenable). It is a reminder that a 2017 LIMS release can still be in service, and that public CVEs exist for this category of software.

Synnovis pathology services, 2024. Synnovis, a pathology services provider in south east London, was hit by ransomware on 3 June 2024. Its services include blood tests, swabs and bowel tests, and the attack led to cancelled appointments and operations at the hospitals it serves. On 20 June the attackers published close to 400 GB of data they said came from Synnovis, reportedly including patient names, health service numbers and descriptions of blood tests (Infosecurity Magazine). It is the clearest recent case of what happens when a lab's core systems go down: results cannot be produced, and the hospitals that depend on them have to triage.

Impact

Loss of availability first, then loss of confidence in data integrity. After an incident, a regulated lab has to show that its records were not altered, which depends on audit trails the attacker may also have reached. Personal data in clinical labs adds breach notification duties.

Detection

  • Log authentication and administrative actions in the LIMS and ship them off the server, so they survive an attack on it.
  • Alert on bulk exports, new admin accounts and changes to audit trail settings.
  • Monitor file shares for mass renames and rewrites, the signature of encryption in progress.

Defence

  • Patch and upgrade on a plan that includes the revalidation effort, rather than deferring until the version is out of support.
  • Put the LIMS behind single sign-on with MFA, and keep its admin interface off general user networks.
  • Test the web application for injection and access control flaws before go-live and after major upgrades, against a test instance you own.
  • Back up the database and file shares offline or immutably, and rehearse a full restore. More on this in our guide to immutable versus offline backups.
  • Write a paper fallback for critical tests, so the lab can keep releasing urgent results while systems are rebuilt.

Attack 3: How does ransomware stop pharma manufacturing?

Ransomware and destructive malware reach plants by spreading from office IT across flat links into operations, or by arriving inside trusted software. The 2017 NotPetya outbreak halted Merck production and cost an estimated 870 million dollars. In 2020 and 2023, Dr. Reddy's and Sun Pharma also reported operational disruption.

Stacked Purdue model levels for a pharma plant from enterprise down to the physical process, with an industrial DMZ and firewalls between each zone, and notes on conduits, remote access and manual fallbacks

Figure 5. Purdue levels used as IEC 62443 zones. Traffic crosses only through named, filtered conduits.

How it works and why

Most plant outages caused by malware never touch a PLC. They happen because the systems that run production (MES, batch management, historians, operator workstations) are Windows machines on networks connected to office IT, and because plants stop when they cannot schedule batches, record them or release them. A plant may shut down deliberately as a precaution once malware is spreading.

Two models help. The Purdue model splits industrial networks into levels, from the physical process at Level 0 up to enterprise systems at Levels 4 and 5 (Wikipedia, Purdue). IEC 62443 is the main series of industrial automation security standards, published with ISA (ISA). It groups assets with common security requirements into zones and connects them through conduits, each with a defined security level (Wikipedia, IEC 62443). Used together, they turn "the plant network" into a set of separately defended areas with controlled crossings. Our guide to IEC 62443 roles, zones and conduits goes deeper.

What the attacker needs

A foothold in office IT, or control of something the company already trusts, such as a software update. Then a flat path from office IT into operations: shared domain accounts, dual-homed machines, broad firewall rules or remote access straight into Level 3.

Real cases

These are described as business cases. None is attributed here to any country or group.

Merck and NotPetya, 2017. On 27 June 2017, NotPetya malware disrupted Merck worldwide. It entered through a server in Merck's Ukraine office that ran a tax and accounting application, M.E.Doc, which had been compromised, and then spread from computer to computer. According to Bloomberg reporting, it crippled more than 30,000 laptops and desktops and 7,500 servers, and hit sales, manufacturing and research. Merck initially estimated the damage at 870 million dollars in regulatory filings, and its production was so disrupted that it could not meet 2017 demand for its Gardasil 9 vaccine (Claims Journal). The business lessons: a trusted update can be the way in, a flat global network spreads the damage, and insurance may be contested for years. More in our NotPetya retrospective.

Dr. Reddy's Laboratories, 2020. In October 2020 the company said that in the wake of a detected cyber attack it had isolated all data centre services as a preventive step, and expected services to be back within 24 hours. Reports said plants in Brazil, India, Russia, the United Kingdom and the United States were affected (Security Affairs). Isolating systems fast limited the damage, at the cost of a short production pause.

Sun Pharmaceutical Industries, 2023. Sun Pharma first reported an IT security incident on 2 March 2023. In a later statement to stock exchanges it said a ransomware group had claimed responsibility, that some file systems had been breached and some company and personal data stolen, and that containment measures had affected operations and would reduce revenue (Insurance Journal).

Impact

Lost production, missed supply commitments and delayed batch release. For vaccines and specialty medicines with few suppliers, that can mean shortages. Every batch made during the incident may need extra quality review, because records were interrupted.

Detection

  • Monitor the boundary between office IT and operations for new connections, especially remote desktop and file sharing protocols crossing it.
  • Use passive OT network monitoring to learn normal traffic between Levels 1 to 3 and alert on new hosts, new protocols and engineering commands outside maintenance windows.
  • Watch privileged accounts used on both sides of the boundary. Shared accounts are how malware crosses.

Defence

  • Zone the plant using the Purdue levels and IEC 62443 zones and conduits, with an industrial DMZ between enterprise and operations. Nothing should cross from Level 4 straight to Levels 1 or 2.
  • Separate identity. Operations should not depend on the corporate domain to log in or run batches.
  • Route all remote access through a DMZ jump host with MFA, approval and session recording.
  • Back up MES, batch, historian and PLC configurations offline, and practise restoring them on spare hardware.
  • Keep manual fallbacks alive. The safety systems that protect people and equipment run independently of IT. Plants can also hold paper batch record templates and documented manual operation procedures, so they can finish batches safely and decide calmly whether to keep running.
  • Treat software updates as supply chain events. Know which vendor update channels reach which zones, and stage updates through the DMZ.

Why are validated GMP systems so hard to patch?

Because a validated system has been formally tested to prove it works as intended, and any change must be assessed, controlled and tested again under the company's quality system. A risk-based approach, as in ISPE GAMP 5, lets teams test only what a patch can affect, which makes security updates practical.

Decision flow for a vulnerability in a validated system: check whether the component is present, whether it is reachable or exposed, and whether a fix exists, leading to a risk-based change or to compensating controls and a minimal backported patch

Figure 6. Prove relevance first, then choose the smallest safe change.

Validation is not the enemy of security, but it does change the economics. A full revalidation of an MES can take weeks of qualified people's time. So the question for each new vulnerability is not "is there a patch" but "is this vulnerability relevant to this system, and what is the smallest change that removes the risk".

A practical path, shown in Figure 6:

  1. Is the vulnerable component actually present? An accurate software bill of materials (SBOM), a machine-readable list of every component and version in a system, answers this in minutes. Without one, teams spend days asking vendors.
  2. Is it reachable or exposed? A flaw in a library function the system never calls, or in a service that is switched off, carries far less risk than one in a network-facing parser. Record the rationale either way; auditors will ask.
  3. Is a fix available? If yes, raise a change, run targeted regression tests on the functions the patch touches, release, and update the SBOM and the validation records. If no, apply compensating controls now (isolation, allow lists, disabled services), and if the code is yours, consider a minimal backported patch that changes as little as possible.

GAMP 5 describes risk-based validation of GxP computerised systems, including the increased role of service providers and software tools (ISPE). Many quality teams already use that risk-based thinking for functional changes. Applying it to security patches, with SBOM and reachability evidence in the change record, is what lets them keep up. Our guide to pharmaceutical software validation in the supply chain covers the documentation side.

Attack 4: How are clinical trial platforms attacked?

Clinical trials depend on a few SaaS vendors for data capture, patient diaries and randomisation, so an attack on one vendor can pause many sponsors' trials at once. In 2020, ransomware at clinical trial software maker ERT forced some trial sites back to pen and paper.

How it works and why

A sponsor running a trial rarely hosts the systems. A contract research organisation (CRO) and a set of SaaS vendors do. That concentrates risk: a single vendor serves hundreds of trials, and its customers inherit its security. Trial data is also sensitive personal health data, and the integrity of the dataset is what regulators review when deciding whether to approve a drug.

What the attacker needs

Whatever gets them into the vendor: a phished employee, an exposed service or a compromised supplier of the vendor. Against the sponsor directly, access to the accounts that integrate with trial systems.

Real case

eResearchTechnology (ERT), 2020. ERT, a Philadelphia company whose software is used in clinical trials, was hit by ransomware on 20 September 2020. The company powered down its systems and brought in outside experts; some systems were offline for around two weeks. Researchers at some sites tracked patients with pen and paper. Customers affected included IQVIA, which was running a COVID-19 vaccine trial, and Bristol Myers Squibb; both said the effect was limited because they had backups. IQVIA said it was not aware of trial data being taken before files were encrypted, and patient safety was not put at risk (HIPAA Journal). ERT software was reported to have been used in about three quarters of the trials that led to drug approvals in the previous year (SiliconANGLE).

Impact

Delayed trials, extra monitoring work, possible data loss for patient diaries, and questions about data integrity that need documented answers. For a sponsor, weeks of delay on a late-stage trial is expensive.

Detection

Sponsors see vendor incidents late. The practical signals are integration failures, unusual data sync gaps and vendor status messages. Contract terms that require prompt incident notice matter more than any tool here.

Defence

  • Assess trial vendors for security as seriously as for quality: SOC 2 or ISO 27001 reports, SBOMs for their platforms, backup and recovery tests, and incident notification terms. Our guide to running a vendor security review for clinical software has a checklist.
  • Hold your own data exports on a schedule, so a vendor outage does not leave you without your data.
  • Agree paper and phone fallbacks for dosing, randomisation and adverse event reporting with your CRO before the trial starts.
  • Limit integration accounts to the scopes they need, with MFA for humans and short-lived credentials for systems.

Attack 5: How does genetic data leak?

Genetic data most often leaks through accounts, not through broken databases. In 2023, attackers logged into about 14,000 23andMe accounts with passwords reused from other breaches, then used a relative-matching feature to reach data on about 6.9 million people. Genetic data cannot be reset after it leaks.

Horizontal bar chart on one scale: about 14,000 accounts accessed directly, about 5.5 million DNA Relatives profiles reached, about 1.4 million Family Tree profiles reached, about 6.9 million people affected in total

Figure 7. One sharing feature turned a small number of logins into millions of affected people.

How it works and why

Credential stuffing means taking username and password pairs leaked from one site and trying them on another. It works because people reuse passwords. On a genomics service, each account also sees information about genetic relatives who opted into matching. So each stuffed account exposes not just one person but many.

What the attacker needs

Lists of leaked credentials, which are widely traded, and a login page without strong rate limiting, breached-password checks or mandatory two-step verification.

Real case

23andMe, 2023. The incident came to light in October 2023 when someone claimed online to hold profile data on millions of users. 23andMe found that its own systems had not been compromised; the attackers had used credential stuffing to access about 14,000 accounts, about 0.1 percent of its roughly 14 million customers. Through the DNA Relatives feature, those accounts gave access to profile data on about 5.5 million people, and family tree data on about 1.4 million more, for about 6.9 million people in total. Data for directly accessed accounts generally included ancestry and some genetic health information; for relatives it included display names, predicted relationships, percentage of DNA shared and sometimes birth year, location and photos. 23andMe reset all passwords and made two-step verification mandatory (HIPAA Journal, TechCrunch).

The story did not end there. 23andMe later went through bankruptcy, and its genetic data became part of the assets for sale. In May 2025 Regeneron Pharmaceuticals won an auction with a 256 million dollar bid; bidding reopened, and on 30 June 2025 a bankruptcy judge approved a sale to TTAM Research Institute, a nonprofit led by a 23andMe co-founder, for 305 million dollars. TTAM pledged to keep allowing customers to delete their data (NPR). The lesson for anyone holding genomic data: a breach is not the only way data changes hands, and customers will judge you on what happens to their data over decades.

Impact

Exposure of ancestry and health-related genetic information, which can reveal family relationships and inherited risks. Unlike a password, a genome cannot be changed, and it implicates relatives who never signed up.

Detection

  • Watch login telemetry for many failed logins across many accounts from rotating addresses, and for successful logins from new devices followed by bulk profile views. Our guide to detecting credential stuffing from application logs shows what to look for.
  • Alert on accounts that view an unusual number of other users' profiles.

Defence

  • Require MFA or passkeys for any account that can see genetic or health data.
  • Check new and existing passwords against known breached passwords, and rate limit logins by account and by source.
  • Design sharing features for abuse. Limit how much a single account can see about others, and how fast.
  • Minimise and segregate. Store genetic data separately from identity data, with separate access controls, and keep only what the stated purpose needs.
  • Plan for the long life of the data, including what happens in a sale or wind-down, and say so in your privacy terms.

Attack 6: How is research and manufacturing IP stolen?

Intellectual property theft in life sciences usually looks like ordinary access used quietly: phished credentials, compromised partner accounts, over-shared cloud folders and insiders taking data when they leave. The targets are research data, assay and process know-how, regulatory dossiers and manufacturing recipes.

How it works and why

A drug programme's value lives in data: compound libraries and assay results, clinical datasets, process development records, batch recipes and the regulatory submission that ties them together. Much of it sits in document systems, ELNs, cloud storage and email, shared with CROs, contract manufacturers and academic partners. Attackers who want it do not need to disrupt anything. They need to stay quiet and read.

The common routes are phishing and credential theft against researchers, compromise of a partner who has legitimate access, cloud storage shared by link or opened to the internet by mistake, and insiders who copy data before moving to a competitor. This guide does not attribute any incident to a country or group; the defences are the same whoever the attacker is.

What the attacker needs

A valid account with read access, and time. Exfiltration is often slow and blends into normal sync traffic.

Impact

Lost competitive position, compromised patent strategy and, for manufacturing know-how, the ability for someone else to copy a process. These losses rarely show up in a quarter's results, which is why they are under-reported.

Detection

  • Data loss prevention and cloud access monitoring tuned for research repositories: large downloads, new sync clients, sharing links to personal accounts.
  • Partner account activity outside agreed projects or hours.
  • Leaver processes that check recent access and downloads before the last day.

Defence

  • Classify the crown jewels (key datasets, process recipes, submission dossiers) and put them in repositories with tighter access and logging.
  • Give partners project-scoped access that expires, not standing access to shared drives.
  • Require MFA and managed devices for anyone who can reach crown jewel data.
  • Review cloud sharing regularly and block public links on sensitive stores.

Attack 7: How do open source packages and Python dependencies compromise research pipelines?

Bioinformatics and AI pipelines install hundreds of open source packages, often unpinned, from public indexes. Attackers exploit that with lookalike or same-name packages and by compromising project build pipelines. The 2022 torchtriton and 2024 Ultralytics incidents both reached users through PyPI, the main Python package index.

Flow diagram of a research pipeline: package indexes, container images, public AI models, reference data and CI workflows feed the environment build, then a workflow engine on cloud compute, then results; red boxes mark dependency confusion, compromised publishing, model files that run code, leaked secrets and open storage

Figure 8. Every input on the left runs with the pipeline's access to data and cloud keys.

How it works and why

When a pipeline runs a package install, it executes code written by strangers, with the pipeline's permissions. Research environments are especially exposed: they favour the latest versions, mix conda and pip, run on machines with access to data and cloud credentials, and are often maintained by scientists rather than build engineers.

Two techniques are documented in Python:

  • Dependency confusion. An organisation uses a package from its own index, and an attacker publishes a package with the same name on the public index. If the installer prefers the public one, the attacker's code runs.
  • Compromised publishing. An attacker gets into a legitimate project's build and release pipeline, for example its CI workflows or a leftover publishing token, and ships malicious versions under the real project's name.

What the attacker needs

For dependency confusion, the name of an internal package and a free account on the public index. For compromised publishing, a weakness in the target project's CI configuration or credential hygiene.

Real cases

PyTorch nightly and torchtriton, December 2022. Users who installed PyTorch nightly builds on Linux through pip between 25 and 30 December 2022 also installed a dependency named torchtriton. A malicious package with that name had been uploaded to PyPI, and because PyPI took precedence over PyTorch's own nightly index, it was installed instead of the real one. Its binary collected system information and read files including the first 1,000 files in the user's home directory, then uploaded them. Stable PyTorch releases were not affected (PyTorch). Our torchtriton explainer covers the mechanics.

Ultralytics, December 2024. The ultralytics package, widely used for computer vision, suffered a supply chain attack through its GitHub Actions workflows and then a PyPI API token. Versions 8.3.41, 8.3.42, 8.3.45 and 8.3.46 were affected and removed. The first malicious releases went out through the project's normal workflow after an attack on the GitHub Actions cache used in the build; the second round used an unrevoked API token. Because the project used Trusted Publishing and publish attestations, investigators could tell which releases came from the real workflow and which did not (PyPI blog, and our write-up on how Trusted Publishing made it auditable).

Neither incident targeted life sciences specifically. Both hit the same ecosystem that drug discovery, imaging and genomics teams use every day.

Impact

Theft of credentials and data from research machines, cloud keys that open storage buckets of sequencing or trial data, and possible tampering with analysis results.

Detection

  • An SBOM for every pipeline environment and container, so you can answer "do we use this package and version" the hour an incident breaks.
  • Alerts on new packages entering lock files, and on packages whose publisher, provenance or install scripts change.
  • Egress monitoring from pipeline compute: analysis jobs rarely need to talk to arbitrary internet hosts.

Defence

  • Pin dependencies with hashes in lock files for conda and pip, and rebuild environments from them. Our comparison of conda lock files and environment files helps here.
  • Use one controlled index or mirror that proxies public packages, so internal names can never be shadowed by public ones.
  • Block known malicious and lookalike packages at install time, not after the fact.
  • Harden your own CI: no shared caches between untrusted and release jobs, short-lived publishing credentials through Trusted Publishing, and pinned actions.
  • Keep secrets out of notebooks and images, and give pipeline jobs least-privilege cloud roles.
  • Know the review model of each ecosystem. Bioconductor, conda-forge, CRAN and PyPI review packages differently; our comparison explains what each check means.

Attack 8: What are the security risks of AI models in drug discovery?

The most concrete risk is model files that run code when loaded: the common pickle format can execute arbitrary code, and researchers found around 100 malicious models on Hugging Face in 2024. Other risks include poisoned training data, leakage of proprietary data through models and tools, and misuse of generative chemistry models.

How it works and why

Many machine learning models are saved with Python's pickle serialisation format. Python's own documentation warns that pickle is not secure, that malicious pickle data can execute arbitrary code during unpickling, and that you should never unpickle data from an untrusted source (Python docs). Loading a downloaded model can therefore be the same as running a downloaded program. Safetensors is a format designed to store model weights safely, as opposed to pickle (Hugging Face).

Beyond file formats:

  • Data poisoning. Manipulating training data so a model learns something wrong, for example to mis-score a class of compounds. This needs access to the training data or its sources.
  • Data leakage. Proprietary structures, sequences or trial data pasted into external AI tools, or memorised by models trained on them and then exposed.
  • Dual use. A 2022 commentary in Nature Machine Intelligence, "Dual use of artificial-intelligence-powered drug discovery", described how AI tools built for drug discovery could be misused to design harmful molecules, and turned that thought experiment into a computational proof, with a model generating molecules predicted to be toxic (Nature Machine Intelligence). For companies, that is a governance question about who can use which models and for what.

What the attacker needs

For malicious model files, only a public upload and a researcher who loads it. For poisoning and leakage, access to data pipelines or to the people and tools that handle data.

Real case

Malicious models on Hugging Face, 2024. In February 2024 JFrog reported that its scanning of Hugging Face had found around 100 models with real harmful payloads, excluding false positives. One PyTorch model, since deleted, opened a reverse shell to a remote host when loaded, giving the uploader control of the machine. The payloads relied on pickle's ability to execute code on load (JFrog). Our deeper guides cover AI model supply chain risks on Hugging Face and pickle deserialisation vulnerabilities in ML frameworks.

Impact

Compromise of research workstations and GPU servers, which usually hold valuable data and cloud credentials. For poisoning, wrong scientific conclusions that may not be noticed for a long time.

Detection

  • Inventory every model in use, where it came from, its format and its hash. A machine learning bill of materials (ML-BOM) records models, datasets and their provenance; CycloneDX supports this as AI/ML-BOM (CycloneDX).
  • Scan model files before use for code execution in pickle and similar formats.
  • Monitor GPU and research servers for unexpected outbound connections.

Defence

  • Load models only in safe formats such as safetensors where possible, and only from sources you have approved.
  • Mirror approved models internally with recorded hashes, rather than pulling from public hubs at run time.
  • Separate experimentation from production, with production models built and signed through a controlled pipeline.
  • Set rules for external AI tools about which data may be used with them.
  • Govern generative chemistry and biology models with access controls and use policies. Our AI governance inventory for a pharma company shows a starting point.

Which attacks matter most for life sciences?

For most organisations the priority order is ransomware through phished or exposed access, credential stuffing on customer accounts, and dependency and model risks in research pipelines, because they are cheap for attackers and costly for victims. Instrument firmware and supplier update attacks are serious but rarer.

Qualitative matrix of attacker effort against business impact, placing ransomware, credential stuffing, dependency confusion and malicious model files in the low effort area, and compromised supplier updates, firmware implants and poisoned training data in the high effort area

Figure 9. An editorial placement to help prioritise. It is not measured data.

The upper left of Figure 9 is where to start: those attacks need little skill and have caused real losses. The upper right, a compromised supplier update, is rare but devastating, as NotPetya showed, and the defence is mostly architectural: segmentation, staged updates and knowing what each update touches. The lower right deserves attention once the basics are in place.

How can you detect an attack in a lab, plant or research pipeline?

Detection in life sciences depends on knowing what normal looks like: an inventory of every instrument, system, pipeline and model, baseline network traffic in labs and plants, centralised logs that survive an attack, and monitoring at the boundaries between office IT, labs, operations and the cloud.

A workable detection programme has five parts:

  1. Inventory with versions. Instruments and their PCs, plant systems, LIMS and ELN, pipeline environments, containers and models, each with an SBOM. You cannot match a new advisory to assets you cannot list.
  2. Boundary monitoring. The crossings between office IT and labs, office IT and the industrial DMZ, and the DMZ and operations. Most serious incidents cross at least one.
  3. Passive OT and lab network monitoring. Specialist platforms such as Claroty, Nozomi Networks and Dragos learn industrial protocols and alert on new devices, new communication paths and engineering actions. They are worth evaluating for plants of any size.
  4. Centralised, protected logs. Authentication, admin actions, LIMS audit events and backup job results, shipped somewhere an attacker on the main network cannot erase.
  5. Supply chain alerts. Notices when a package, model, container base image or vendor component you use gets a new vulnerability or a malicious version.

Detection only helps if someone acts. Write playbooks for the scenarios in this guide, including who can decide to isolate a plant or a lab network, and rehearse them.

How do the defences fit together?

Defence in life sciences works in layers: inventory and SBOMs, strong identity, network zoning, hardened endpoints and instruments, a controlled software and model supply chain, protected data and backups, and practised detection and response with manual fallbacks. Each layer limits the damage when the one above it fails.

Seven stacked defence layers from inventory and SBOMs through identity, network, endpoints and instruments, software and model supply chain, data, and detection and response

Figure 10. Seven layers. Each one limits the damage when the one above it fails.

Two of these layers carry more weight in life sciences than elsewhere. Network zoning matters because so many systems cannot be patched quickly; segmentation buys the time validation needs. Manual fallbacks matter because plants, labs and trials have regulated processes that can continue on paper if they are planned for, which turns an outage into a slowdown.

Hardening checklists for biotech labs, pharma manufacturers and instrument makers

Labs should segment instruments, remove shared logins and pin pipeline dependencies. Manufacturers should zone plants with IEC 62443, control vendor access and use risk-based patching. Instrument and lab software makers should ship SBOMs, enable Secure Boot, authenticate every service and offer a supported upgrade path.

Three-column checklist for biotech and research labs, pharma manufacturers, and instrument and lab software makers, six items each

Figure 11. Start at the top of your column. Each item closes a door used in a documented case.

For biotech companies and research labs

  • Inventory every instrument with its PC, operating system, vendor software versions and exposed services.
  • Put instruments on their own network segment with no direct internet access and only the connections each needs.
  • Remove shared instrument logins and local administrator rights for everyday users; use named accounts where the vendor allows.
  • Route vendor remote support through a jump host with MFA, enabled per session.
  • Apply vendor security updates through the vendor's process, and plan operating system upgrades before end of support.
  • Protect the LIMS and ELN with single sign-on and MFA, and keep their admin interfaces off user networks.
  • Pin pipeline dependencies with hashes, install from one controlled mirror, and keep an SBOM for each environment and container.
  • Load AI models only in safe formats from approved sources, and record their hashes.
  • Keep secrets out of notebooks, scripts and images; use short-lived cloud credentials for jobs.
  • Back up raw instrument data, LIMS databases and pipeline outputs offline or immutably, and test restores.
  • For consumer or patient-facing genomics, require MFA, check passwords against breach lists and rate limit sharing features.

For pharma and biologics manufacturers

  • Map each plant to Purdue levels and define IEC 62443 zones and conduits, with an industrial DMZ.
  • Give operations its own identity domain; no shared accounts across the IT and OT boundary.
  • Allow vendor and engineer remote access only through the DMZ, with MFA, approval and session recording.
  • Use a documented risk-based change process for security patches, with SBOM and reachability evidence in the change record.
  • Deploy passive monitoring on OT networks and alert on new hosts, protocols and engineering commands.
  • Back up MES, batch, historian, HMI and PLC configurations offline, and rehearse rebuilding them.
  • Maintain and rehearse manual operation and paper batch record procedures for critical products.
  • Require SBOMs, vulnerability disclosure policies and supported lifecycles from equipment and software suppliers.
  • Stage and verify vendor software updates in the DMZ before they reach operations.
  • Run incident exercises that include quality, production and supply chain leaders, not only IT.

For lab instrument and lab software makers

  • Ship a complete SBOM (CycloneDX or SPDX) with every software release and every instrument PC image, including the board firmware and its supplier.
  • Enable Secure Boot and firmware write protection, and avoid legacy CSM boot on new designs.
  • Bind each network service only to the interfaces it needs, require authentication by default, encrypt traffic and run services without administrator rights.
  • Sign software and firmware updates, and give customers a supported path off end-of-life operating systems before support ends.
  • Publish a vulnerability disclosure policy and security advisories with affected versions and mitigations.
  • Run SAST, SCA, secrets scanning and DAST in your pipeline before each release, and track vulnerabilities in the components you ship.
  • Design remote support so the customer controls when it is on, with MFA and logging.
  • Give customers the documentation they need for risk-based validation of your updates: what changed, what to retest and why.
  • Ask your board, operating system and SDK suppliers for SBOMs and update commitments, and hold them to it.

Where does the life sciences software supply chain break?

It breaks wherever a company runs code it cannot see: board firmware inside instruments, vendor services on instrument PCs, LIMS and MES platforms with their libraries, SaaS trial platforms, open source packages in pipelines and public AI models. SBOMs, VEX statements and ML-BOMs are how each link becomes visible.

The cases in this guide map onto the chain:

  • Board and firmware suppliers. The iSeq 100's BIOS came from its board supplier, and the researchers expected similar issues in other devices using the same boards. Instrument makers inherit their suppliers' firmware security. See our firmware supply chain security guide.
  • Instrument software vendors. LRM and UCS were vendor services shipped on the instrument. Labs depend on the vendor to find and fix such flaws, and on an SBOM to know which instruments are affected.
  • Lab and plant platforms. LIMS, ELN, CDS and MES products carry their own open source libraries, which are rarely visible to customers without an SBOM.
  • Trusted updates. NotPetya arrived through a compromised update for business software. Any update channel that reaches operations is a supply chain link.
  • SaaS platforms. The ERT incident showed concentration risk across sponsors.
  • Open source and models. torchtriton, Ultralytics and the Hugging Face models showed how public ecosystems deliver malicious code directly into research environments.

What makes these links visible:

  • SBOMs in CycloneDX or SPDX for every product release, instrument image, plant system and pipeline environment, delivered by suppliers and generated internally.
  • VEX (Vulnerability Exploitability eXchange) statements, which say whether a known vulnerability actually affects a product, so customers do not chase irrelevant CVEs. That is exactly the evidence a risk-based change process needs.
  • ML-BOMs for models and datasets, recording provenance and formats. Our guide asks whether you need an ML-BOM or an SBOM is enough.
  • Supplier requirements in purchase contracts: SBOM delivery, disclosure policies, update commitments and supported operating system roadmaps.

For instrument and software makers selling into GxP customers, SBOMs and vulnerability handling records are also becoming part of supplier audits. Our guide to passing a pharma customer's GxP supplier audit covers what auditors ask for.

How Safeguard helps life sciences companies and instrument makers

Safeguard covers a life sciences software supply chain in one platform: SBOMs for source, containers, binaries and disk images, reachability-ranked vulnerabilities, SCA, SAST, secrets scanning and DAST for code you own, model scanning, fix pull requests gated by your CI, Gold patched packages, supplier risk and compliance evidence, deployed as SaaS, in your VPC or on premises.

Flow diagram: source repositories, container images, binaries and disk images, and AI models feed Safeguard's SBOM and vulnerability step, then reachability ranking, then fixes through CI-tested pull requests and Gold patched packages, producing SBOM and VEX, compliance evidence and policy gates

Figure 12. One inventory from source to shipped image, ending in a fix and the evidence for it.

A life sciences organisation typically has several codebases: in-house pipelines and analysis code, LIMS and MES extensions and integrations, cloud services for partners and patients, and, at instrument and lab software makers, the instrument software and PC images themselves. Teams often end up with one tool per codebase, plus an SBOM manager, a vulnerability tracker and a governance, risk and compliance (GRC) tool for audits, each with its own inventory. Safeguard does that work end to end with one inventory and one vulnerability model, from the first SBOM to the merged fix and the evidence an auditor asks for.

See inside binaries and images you ship and receive

In the Safeguard console, Integrations > Upload App Binary accepts Windows EXE, MSI and DLL files, Android APKs, iOS IPAs, VM and disk images (QCOW2, VMDK, VHD and OVA) and container tarballs. Safeguard unpacks the file, identifies embedded libraries and their versions, maps them to known vulnerabilities with EPSS and known exploited enrichment, records licences and creates a project with a full SBOM.

For life sciences, that covers instrument control software installers, a disk image of an instrument PC or a plant workstation, vendor-supplied LIMS components, and the containers your pipelines run in. Raw firmware images, such as a BIOS dump, may need to be extracted into their component files first; Safeguard does not claim to parse every raw firmware format directly. Heavy obfuscation, encryption or unusual compression can give incomplete results. More in our guide to scanning app binaries with Safeguard.

Scan the code you own

For pipelines, LIMS extensions, instrument software and cloud services in your repositories, Safeguard runs software composition analysis (SCA) across the full transitive dependency tree of the languages your pipelines and products use. Static application security testing (SAST) checks your own code for bug classes such as injection and missing authentication, the classes behind the LIMS and instrument service CVEs above. Secrets scanning catches cloud keys and tokens in code and notebooks before they ship. Dynamic application security testing (DAST) tests the web applications and APIs you run, against verified targets you own.

At install time, Safeguard's Package Firewall sits in front of npm and pip and can block known malicious packages, typosquats and dependency confusion attempts, which is the torchtriton pattern.

Scan the models you load

Safeguard's model scanning inspects model artifacts, including pickle, safetensors, GGUF and ONNX files, for embedded code and malware, and AI-BOM discovery inventories the models and AI SDKs your code uses. That gives research teams an approved model list with evidence behind it.

Rank by reachability, not by count

A pipeline environment or instrument image can carry hundreds of known vulnerabilities in its libraries. Reachability analysis shows which ones your code can actually reach, so a flaw in a network-facing parser rises above one in an unused function. For a validated system, that is the evidence step 2 of Figure 6 needs, recorded in a form an auditor can follow. Our guide on reachability for research code with many dependencies shows how it plays out.

SBOMs customers and auditors can use

Every project produces an SBOM you can export in CycloneDX or SPDX, among many other formats, for every release, image and environment. Attestation and signing with SLSA, Sigstore and in-toto record build provenance. CSAF and VEX documents can be uploaded and linked, so the SBOM an instrument maker hands a pharma customer comes with a clear statement of which vulnerabilities actually affect it.

Fix it without waiting for upstream

This is where Safeguard differs most from a scanner. 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, and your policies decide what can merge. In a regulated environment that matters: the fix enters your normal change control and verification process as a reviewed, tested change, not an untracked edit.

For open source components where upstream has not shipped a fix, or where the fix is in a major version you cannot move to on a validated system, Safeguard Gold provides patched, hardened builds of those packages, and Safeguard can fork and patch the vulnerable package so your next build pulls a fixed version with the smallest possible change. A minimal backported fix is far easier to assess and retest under a risk-based change than a major upgrade, which is the right-hand path of Figure 6.

When the vulnerable code sits inside a supplier's binary, such as an instrument vendor's service or a board supplier's firmware, Safeguard tells you exactly which products, images and sites contain it. That lets you press the supplier with evidence, apply compensating controls where they matter, and give quality teams an accurate risk statement while you wait.

Supplier risk and compliance evidence

Safeguard's Third-Party Risk Manager lets you request SBOMs from instrument makers, LIMS and MES vendors, trial platform providers and contract manufacturers, score them and monitor them continuously. The compliance suite maps evidence from the same scans to frameworks and regulations such as ISO 27001, SOC 2, HIPAA and GDPR, so SBOMs, vulnerability handling records and fix history support audits and the security questionnaires pharma customers send. Guard adds runtime guardrails on Linux servers, inspecting system calls, library calls and network egress.

Deploy where your data lives

Research data, trial data and process know-how are sensitive. Safeguard runs as SaaS in US and EU regions, in your own VPC or fully on premises, and can run in Zero AI mode with no language model in the loop, or with your own model. The local runner CLI scans on your own machines and sends only encrypted results. Safeguard is SOC 2 Type I and Type II audited.

How this compares

Life sciences teams often combine a source SCA tool such as Snyk or Black Duck, a model scanning tool, an OT visibility platform such as Claroty, Nozomi Networks or Dragos, and a GRC platform. Those are capable products, and OT platforms go deeper on industrial protocol monitoring than Safeguard does; JFrog also produced the model research cited here. Safeguard's case is consolidation and remediation across the software supply chain: one inventory across source, containers, binaries, images and models; reachability-ranked vulnerabilities; fix pull requests and patched packages instead of tickets; and supplier risk and compliance evidence from the same data. Safeguard can sit alongside an OT monitoring platform and take over the SBOM, fixing and evidence work.

The takeaway

The documented attacks on life sciences are mostly ordinary attacks on unusual systems. Ransomware crossed from office IT into plants and trial platforms. Reused passwords opened genetic data. Malicious packages and model files reached research machines. Instrument flaws, including ones rated CVSS 10, were found and fixed by vendors before any reported exploitation.

What makes the sector hard is that its systems are validated, long-lived and full of supplier code, and its data never stops being sensitive. The answer is the same for a ten-person biotech and a global manufacturer: know every component in every system, separate labs and plants from office IT, keep manual fallbacks alive, and fix through a risk-based path that quality teams trust. Instrument and software makers can make that far easier by shipping SBOMs, secure defaults and timely fixes.

Try it yourself

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

Test yourself

Why are validated GMP systems so hard to patch?

A validated system has been formally tested to prove it works as intended, and any change must be assessed, controlled and tested again under the quality system. A risk-based approach such as ISPE GAMP 5 lets teams test only what a patch can affect, which makes security updates practical.

What went wrong with the iSeq 100 sequencer firmware that Eclypsium analysed in 2025?

It booted a 2018 BIOS build with known vulnerabilities, ran in Compatibility Support Mode, had firmware read and write protections disabled and did not use Secure Boot. An attacker with access could overwrite the firmware to brick the instrument or plant a persistent implant.

How did the 2023 23andMe incident reach about 6.9 million people?

Attackers used credential stuffing, logging into about 14,000 accounts with passwords reused from other breaches. The DNA Relatives and family tree features then gave access to profile data on about 5.5 million and 1.4 million more people. Genetic data cannot be reset once it leaks.

Why did the 2020 ransomware attack on clinical trial software maker ERT affect so many sponsors?

Trial data capture and diaries run on a few SaaS vendors, so one vendor serves many trials. Some ERT systems were offline for about two weeks and some sites went back to pen and paper. Customers such as IQVIA and Bristol Myers Squibb said the effect was limited because they had backups.

What happened to Merck in the 2017 NotPetya outbreak, and how did it get in?

NotPetya entered through a compromised tax and accounting application, M.E.Doc, on a server in Merck's Ukraine office, then spread from computer to computer. Merck estimated the damage at 870 million dollars, and production was disrupted enough that it could not meet 2017 demand for its Gardasil 9 vaccine.

Why can loading a downloaded AI model be risky, and what is a safer format?

Many models are saved in Python's pickle format, and Python's own documentation warns that malicious pickle data can execute arbitrary code when loaded. Safetensors is a format designed to store model weights safely, as opposed to pickle.

Frequently asked questions

Can a DNA sequencer be hacked?

Researchers and vendors have documented serious vulnerabilities in sequencer software and firmware. Illumina's Local Run Manager (2022) and Universal Copy Service (2023) had flaws rated CVSS 10, and a 2025 study found an iSeq 100 booting a 2018 BIOS without Secure Boot. Fixes were released and no exploitation was reported. Keep instruments patched and on isolated networks.

Why do lab instruments still run Windows 7?

Because the instrument and its PC were tested and released as one unit, and the vendor controls the upgrade. Illumina, for example, says instruments keep working on Windows 7 but that an unsupported operating system raises malware risk, and it arranges Windows 10 upgrades through technical support. Until upgraded, isolate the instrument and limit its connections.

Can you patch a GxP validated system without revalidating everything?

Usually yes. A risk-based approach, as described in ISPE GAMP 5, lets you assess what a patch can affect and test those functions rather than revalidating the whole system. SBOM evidence that a component is present, and reachability evidence that it matters, make the change record stronger and the decision faster.

What happened to Merck in the NotPetya attack?

In June 2017 NotPetya malware entered Merck through a compromised accounting application in its Ukraine office and spread across its network, crippling tens of thousands of computers. Production was disrupted badly enough that Merck could not meet 2017 Gardasil 9 demand. Merck initially estimated damage at 870 million dollars and settled an insurance dispute in 2024.

How did the 23andMe breach happen?

Attackers used credential stuffing, logging into about 14,000 accounts with passwords reused from other breaches. 23andMe said its systems were not compromised. Through the DNA Relatives feature, those accounts exposed data on about 6.9 million people. The company reset all passwords and made two-step verification mandatory.

Are AI models used in drug discovery a security risk?

Yes, in specific ways. Models saved in Python's pickle format can run code when loaded, and researchers found around 100 malicious models on Hugging Face in 2024. Teams should load models in safe formats such as safetensors, from approved sources, and keep an ML-BOM of every model and dataset they use.

How should a pharma plant network be segmented?

Use the Purdue model levels as IEC 62443 zones: enterprise systems, an industrial DMZ, site operations such as MES, supervisory control, basic control and the physical process. Connect zones only through filtered conduits, route remote access through the DMZ with MFA, and never let traffic cross straight from the enterprise to controllers.

What should a lab instrument SBOM include?

Every software component on the instrument and its PC: the control and analysis software, vendor services, the operating system and its version, third-party libraries, and the board's BIOS or UEFI firmware with its supplier and version. It should be machine readable (CycloneDX or SPDX), updated with each release, and paired with VEX statements.

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.