How Quantum Computers Are Cooled: Dilution Fridges and Helium-3

Superconducting and most spin qubit quantum computers are cooled by a dilution refrigerator, a cryostat that uses a mixture of helium-3 and helium-4 to hold the processor at roughly 10 millikelvin, about a hundredth of a degree above absolute zero. A pulse tube cooler takes the system from room temperature to about 3 K, and the helium-3 circulation does the rest. Trapped ion and neutral atom machines do not need this refrigerator for their qubits.

The hard part is not reaching the temperature. It is staying there while hundreds of coaxial cables carry control pulses down and readout signals back up, because every cable conducts heat and every attenuator turns signal power into more heat. This post covers the physics, the stages from the 50 K plate to the mixing chamber, why helium-3 is scarce, and how a wiring heat budget caps qubit count. It goes one level deeper than our explainer on what quantum computing is and how qubits and gates work.

Key takeaways

• Superconducting qubits work at a few gigahertz. At 300 K a 5 GHz mode holds about 1,250 thermal photons on average; at 20 mK it holds about six in a million. That ratio is the main reason for the millikelvin requirement.
• A dilution refrigerator cools by moving helium-3 atoms across a phase boundary in a helium-3 and helium-4 mixture, a process that keeps working down to a few millikelvin.
• A modern dry system stacks five plates: about 50 K and about 4 K from a pulse tube cooler, then the still near 0.9 K, the cold plate near 0.1 K and the mixing chamber at about 6 to 20 mK. Cooling power falls from tens of watts to tens of microwatts.
• Helium-3 comes almost entirely from the decay of tritium in nuclear weapons programs. A Congressional Research Service report put new US supply at about 8,000 liters a year against demand of about 80,000 liters in 2008.
• Wiring is the scaling wall. In one published 100-qubit-scale setup, each stainless steel drive line added about 1 mW at the 4 K stage, against 1.5 W of cooling there.
• Trapped ion and neutral atom qubits live in ultra-high vacuum and are laser cooled. Some ion traps use a 4 K cryostat for better vacuum, but none needs a mixing chamber.

Why do superconducting and spin qubits need millikelvin temperatures?

Two physical reasons push superconducting qubits toward the bottom of the temperature scale: thermal noise in the microwave band they use, and the superconductor itself.

Thermal photons at gigahertz frequencies

A superconducting qubit is an electrical circuit, a small aluminium or niobium structure with one or more Josephson junctions, whose two lowest energy levels are separated by a transition frequency typically between 4 and 8 GHz. Readout uses a microwave resonator at a similar frequency. Any electromagnetic mode at that frequency, in the chip, the resonator or the cables, holds thermal photons set by its temperature, and those photons can flip the qubit, dephase it or shift the readout.

The average number of thermal photons in a mode is the Bose Einstein occupation number:

n(T, f) = 1 / (exp(h f / (k T)) - 1)

Here h is the Planck constant and k is the Boltzmann constant. Since the 2019 SI redefinition both are exact: the BIPM defines the kelvin by fixing k at 1.380 649 x 10^-23 J/K, and NIST's CODATA tables list h as 6.626 070 15 x 10^-34 J s. For a 5 GHz mode that gives a characteristic temperature h f / k of 0.240 K.

Worked example 1: thermal photons in a 5 GHz mode. These are direct calculations from the formula; the only choice is the frequency.

Above about 0.24 K the mode holds one or more thermal photons on average and a qubit in it would be randomly excited. Below about 50 mK the occupation drops exponentially and the qubit relaxes to its ground state and stays there. A 4 K liquid helium bath, the classic low temperature environment, still leaves about 16 photons per mode, far too many.

The same formula explains why the wiring needs attenuators: a coaxial cable that starts at room temperature carries about 1,250 photons per mode of 300 K noise toward the chip unless something absorbs it on the way.

The superconducting gap and quasiparticles

In a superconductor, electrons are bound into Cooper pairs, separated by an energy gap from single particle excitations called quasiparticles. Thermal energy breaks pairs, and the equilibrium quasiparticle population falls exponentially with temperature. Quasiparticles tunnelling across a Josephson junction can flip or dephase the qubit. Aluminium, the usual junction material, superconducts only below roughly 1.2 K, so running at 10 to 20 mK puts it far below its transition and makes thermal quasiparticles negligible. The remaining ones come from stray radiation, which is why the chip sits inside infrared shielding.

Spin qubits: same physics, a little more headroom

Spin qubits in silicon or germanium quantum dots store information in the spin of one electron or hole. Their splitting is set by a magnetic field and is often well above 5 GHz, which gives more headroom, but their charge sensors and electron reservoirs also need cold electrons, so most experiments run at tens of millikelvin.

That is starting to change. In a paper posted in February 2019 and published in Nature in 2020, Yang and colleagues reported a two qubit silicon unit cell operated at about 1.5 K, with single qubit gate fidelities of 98.6 percent. They note that dilution refrigerator cooling power is "severely limited" below 100 mK, so running hotter gives far more cooling to spend on control electronics. It is a research result, not a commercial machine, but it shows the millikelvin requirement is a design consequence, not a law.

Why trapped ions and neutral atoms do not need a dilution refrigerator

Trapped ion and neutral atom qubits are individual atoms held in vacuum by electric fields (ions) or focused laser beams called optical tweezers (neutral atoms). Their qubit states are internal atomic levels, isolated from the warm walls by vacuum rather than refrigeration. The atoms are cooled to microkelvin motional temperatures with lasers, not with a refrigerator. NIST's Ion Storage Group describes its work this way: ions "confined in vacuum in electromagnetic traps", cooled and manipulated with lasers and fields.

The neutral atom logical processor reported by Bluvstein and colleagues in Nature in December 2023, with up to 280 physical qubits, used reconfigurable atom arrays and laser control; no dilution refrigerator appears in the qubit path.

Some trapped ion systems do use a cryostat, for vacuum rather than thermal noise. A 2016 paper from the University of Innsbruck describing a cryogenic setup for trapped ion quantum computing explains that background gas collisions knock ions out of the trap, and that a cryostat at 4 K reaches vacuum pressures about four orders of magnitude lower than typical room temperature chambers.

How does a dilution refrigerator work?

A dilution refrigerator cools by moving helium-3 atoms from a concentrated liquid phase into a dilute liquid phase. That transfer absorbs heat much as evaporation does, but it keeps working at temperatures where ordinary evaporative cooling has stopped.

Why evaporation alone is not enough

Pumping on liquid helium cools it by evaporation. Helium-4 gets to about 1 K this way. Helium-3, which has a higher vapour pressure, gets lower, practically to about 300 mK according to a 2025 paper by Kim and colleagues at a cryostat maker. Below that, the vapour pressure falls so steeply that no pump removes atoms fast enough to give useful cooling.

Phase separation in the helium-3 and helium-4 mixture

Helium-4 atoms are bosons and become superfluid at low temperature. Helium-3 atoms are fermions. Mixed together, they stay fully miscible until the mixture is cooled below a critical point. The Kim paper gives that point as 872 mK at a concentration of 67.3 percent helium-3. Below it, the liquid separates into two layers:

• A concentrated phase that is nearly pure helium-3. It is lighter and floats on top.
• A dilute phase of mostly superfluid helium-4 with some helium-3 dissolved in it.

The key property is that the dilute phase keeps a finite helium-3 concentration, about 6.6 percent in standard references, even as the temperature approaches absolute zero. The superfluid helium-4 barely interacts with the dissolved helium-3, so the dilute helium-3 behaves like a gas and the concentrated phase like a liquid. Moving helium-3 from the "liquid" into the "gas" absorbs heat. Because the "gas" density does not vanish at low temperature, the cooling does not die off.

The circulation loop

The phase boundary sits in the mixing chamber, the coldest point of the refrigerator and the plate where the quantum processor is mounted. To keep cooling, helium-3 must keep crossing that boundary, so the machine runs a closed loop:

1. Still. The dilute phase connects up to the still, held at roughly 0.7 to 0.9 K. There helium-3 has a far higher vapour pressure than helium-4, so pumping removes mostly helium-3. A heater on the still sets the circulation rate.
2. Room temperature pumps. The helium-3 rich vapour leaves the cryostat to turbomolecular and backing pumps, then passes cold traps that remove contaminants.
3. Return and condensing. The gas comes back down, is cooled at the 50 K and 4 K stages and condensed into liquid.
4. Heat exchangers. The returning helium-3 passes through counterflow heat exchangers, often sintered silver powder with an enormous surface area, cooled by the cold dilute stream heading up.
5. Back into the mixing chamber. The cold helium-3 joins the concentrated phase and crosses the boundary again, absorbing heat from the mixing chamber and everything bolted to it.

Cooling power at the mixing chamber depends on the helium-3 flow and falls steeply with temperature. Krinner and colleagues at ETH Zurich measured about 5 µW at 13 mK and about 13 µW at 20 mK at one fixed circulation rate, and 19 µW at 20 mK with more still heating. Going from 20 mK to 13 mK cost more than half the cooling.

Wet versus dry systems

Older "wet" dilution refrigerators sit in a dewar of liquid helium-4, refilled regularly. Most systems built for quantum computing are "dry": a pulse tube cryocooler provides the 50 K and 4 K stages, and the only helium is the sealed mixture circuit plus the pulse tube's own working gas. A 2014 paper by Uhlig estimated that over 100 new commercial dilution refrigerators were being built each year and noted that dry designs free up space around the mixing chamber, with mounting plates up to about 50 cm across. That space is what makes large wiring harnesses possible.

What are the stages from room temperature to the mixing chamber?

A dry dilution refrigerator is a stack of copper plates hung from a room temperature top flange, each colder than the one above, each inside its own radiation shield, all inside a vacuum can. The 2019 EPJ Quantum Technology paper by Krinner and colleagues, "Engineering cryogenic setups for 100-qubit scale superconducting circuit systems", published the measured temperatures and cooling powers of the commercial system they used. They describe the values as typical for state of the art pulse tube based dilution refrigerators.

Source: Krinner et al., EPJ Quantum Technology, 2019, Table 1. Temperatures are for the empty refrigerator; cooling powers are at the temperatures shown.

From top to bottom, cooling power falls from 30 W to 19 µW, a factor of about one and a half million. The heat each plate can absorb shrinks far faster than its temperature, and that drives every wiring decision.

Why each stage exists

• 50 K plate. Intercepts room temperature radiation and absorbs most of the heat conducted down cables and supports. It has watts to spare, so this is where you want heat to land.
• 4 K plate. Hosts the first serious attenuators and the high electron mobility transistor (HEMT) amplifiers on the output lines, which in the ETH Zurich setup give about 40 dB of gain.
• Still. About 0.9 K with tens of milliwatts of cooling, useful for anchoring cables before heat reaches lower stages.
• Cold plate. Around 100 mK, a staging point where more attenuators thermalise cable conductors.
• Mixing chamber. The qubit plate, where everything is counted in microwatts.

The pulse tube cooler and vibration

The pulse tube cryocooler that provides the top two stages has no moving parts at the cold end. A compressor drives an oscillating helium gas flow through a valve, so the cold head vibrates less than older coolers with moving displacers, but it still vibrates. Schmoranzer and colleagues measured accelerations from 0 to 50 kHz at the mixing chamber plate and top flange of a cryogen free dilution refrigerator in 2018. They found the dominant contribution, "surprisingly", near 20 kHz, and tested how frame and compressor coupling changed it. Vibration matters because it heats the mixing chamber and generates noise in cables.

The thermodynamic price of cold

The Carnot limit on the work needed to pump 1 W of heat from temperature T up to 300 K is (300 minus T) divided by T.

Worked example 2: minimum work per watt removed. This is a theoretical floor; real cryocoolers do much worse.

So even in an ideal machine, a microwatt at the mixing chamber costs as much work as about 15 mW of heat at room temperature. That is why engineers chase every nanowatt reaching the bottom plate.

Why is helium-3 scarce, and what does it cost?

Helium-3 is a stable isotope with two protons and one neutron. CRS puts its natural abundance at only about 1.37 parts per million of all helium, far too little to separate economically at scale. Almost all helium-3 comes instead from the radioactive decay of tritium.

Where helium-3 comes from

Tritium decays into helium-3 with a half life of about 12.3 years. Congressional Research Service report R41419, "The Helium-3 Shortage: Supply, Demand, and Options for Congress" (versions dated September 2010 to October 2011), explains that the main US source is the nuclear weapons program. The National Nuclear Security Administration produces tritium for warheads; as it decays into helium-3 the mixture is recycled and the helium-3 is separated as a by-product.

CRS also notes that the US produced tritium in heavy water reactors at the Savannah River Site until the last one, the K reactor, shut down in 1988, and that it now irradiates lithium in a light water reactor. Because only about 5.5 percent of a tritium stock decays each year, new tritium production takes years to become a meaningful helium-3 supply.

Worked example 3: the decay arithmetic. With a half life of 12.32 years, the fraction of tritium left after one year is 2 to the power of minus 1/12.32, about 0.945, so about 5.5 percent decays each year, matching CRS. Take an illustrative tritium stock that would yield 1,000 liters of helium-3 gas if it decayed completely: the first year gives about 55 liters. Doubling tritium production adds only about 55 more liters a year, not a doubling of supply overnight. (Gas volumes like these are quoted at standard conditions; our ideal gas law explainer shows how to convert them to moles.)

The shortage

Supply outran demand for years and built a stockpile. After September 2001 the federal government deployed helium-3 neutron detectors at US borders, and demand rose sharply. CRS states that demand was about 8,000 liters a year before 2001 and reached about 80,000 liters in 2008, while tritium decay produced about 8,000 liters of new helium-3 a year. CRS cites the Government Accountability Office as identifying weaknesses in DOE's management of the stockpile as key factors in the slow response.

CRS also notes that "some types of cryogenic research can only be accomplished using helium-3", whereas neutron detection and medical imaging have alternatives. A dilution refrigerator has no substitute working fluid.

What does it cost?

Treat every figure here as historical. CRS reported historical auction prices of about $40 to $85 per liter of gas, market prices historically of $100 to $200 per liter, and commercial prices of $2,000 per liter or more after the shortage became clear. It also reproduced GAO figures for government allocated prices:

Source: CRS R41419, Table 2, citing GAO. These are 2009 to 2011 figures, not current prices. Today the US Department of Energy Isotope Program, through the National Isotope Development Center at isotopes.gov, runs the government's isotope sales; ask it or a gas supplier for a current quote.

The mixture is a closed loop, so helium-3 is bought once and kept. In practice that means leaks are expensive, the mixture sits in a storage tank at room temperature when the system is warm, and the gas handling is designed never to vent helium-3, a point the 2025 Kim paper also makes.

How do signals get in and out?

A superconducting processor needs three kinds of connection to room temperature electronics: microwave drive lines for gates, flux lines for tuning frequencies and two qubit gates, and output lines carrying readout signals back up.

Drive and flux lines: attenuation to remove thermal noise

From worked example 1, the inner conductor of a coax that starts at 300 K carries about 1,250 thermal photons per mode at 5 GHz. Attenuators fix that. Each is a resistive element anchored to a stage. It absorbs a fraction of the incoming signal and noise, turns it into heat at that stage, and emits only its own stage's thermal noise. Stacked down the chain, they replace 300 K noise with 4 K noise, then 100 mK noise, then mixing chamber noise.

Krinner and colleagues calculated that at least about 60 dB of total attenuation, a power factor of one million, is needed to bring room temperature noise down to a thermal photon occupation of about 10^-3 at the mixing chamber. For drive lines they put 20 dB at the 4 K stage, 20 dB at the cold plate and 20 dB at the mixing chamber, giving a noise photon number of about 0.1 percent. Flux lines, which carry slower signals, use less attenuation plus low pass filters.

Worked example 4: attenuation arithmetic. At 300 K and 6 GHz the occupation number is about 1,040. Dividing by 10^6, the effect of 60 dB, gives about 0.001, the target in the paper. The paper assumes an average drive power of minus 78 dBm (about 16 picowatts) at the mixing chamber, so the room temperature source must supply roughly minus 18 dBm (about 16 microwatts) plus cable losses. Almost all of that power becomes heat inside the refrigerator, and the last 20 dB attenuator turns 99 percent of what reaches it into heat at the coldest plate. That is why most attenuation sits higher up, and why active load from simultaneous gate pulses must be budgeted alongside passive conduction.

Output lines: circulators, amplifiers and superconducting cable

Readout runs the other way and the signal is a few photons, so it cannot be attenuated. The qubit must still be protected from noise coming down from warm amplifiers. In the ETH Zurich design:

• Josephson travelling wave parametric amplifiers at the mixing chamber add 20 to 30 dB of near quantum limited gain across 3 to 12 GHz.
• Circulators and isolators pass the signal upward while sending anything travelling downward into a cold matched load, giving more than 60 dB of isolation toward the sample.
• Niobium titanium coax, superconducting and with low thermal conductivity, carries the signal from the mixing chamber to 4 K with almost no loss. The measured passive load per line was about 0.29 µW at the cold plate and about 0.020 µW at the mixing chamber.
• HEMT amplifiers at 4 K add about 40 dB of gain before the signal climbs stainless steel coax to room temperature.

Spin qubit setups add many DC lines for gate voltages and bias currents: twisted pairs of resistive wire heat sunk at every stage and filtered near the mixing chamber.

Shielding around the chip

• Magnetic shielding. High permeability and superconducting cans cut stray fields that trap flux vortices or shift qubit frequencies.
• Infrared shielding. A light tight can with absorptive coatings stops stray infrared photons that break Cooper pairs.
• Radiation shields at each stage block thermal radiation from warmer surfaces.

Why wiring heat load limits qubit count

Every cable is a thermal bridge between stages. After installing 66 RF lines, mostly 0.085 inch stainless steel coax, Krinner and colleagues saw the mixing chamber base temperature rise from 6.1 mK to 8.4 mK, a total passive load of 1.4 µW. That was over ten times below the cooling available at 20 mK, enough for "tens to hundreds" of qubits, but it shows the slope. Their measured per line loads are among the clearest public numbers for this cable type.

Source: Krinner et al., 2019, Table 2. Each value has a stated uncertainty; for example the 50 K drive line value is 45 plus or minus 23 mW.

Worked example: how many coax lines can a 4 K stage carry?

Worked example 5: passive heat load at the 4 K stage. This combines the measured per line figures above with illustrative line counts and an illustrative HEMT dissipation. It is an illustration, not a specification of any machine.

Assume each qubit needs one drive line (1.0 mW at 4 K) and one flux line (1.2 mW), so 2.2 mW per qubit, and that eight qubits share one readout line whose HEMT amplifier dissipates an illustrative 5 mW at 4 K. Available cooling is 1.5 W at 4.2 K.

The table ignores active loads from signals, which add more. Now check the other stages. At 50 K each qubit's two lines add about 100 mW against 30 W, so that stage fills at about 300 qubits. At the mixing chamber, about 0.04 µW per qubit against 19 µW at 20 mK allows a few hundred qubits before counting dissipation in the final attenuators.

With this cable type and this refrigerator, a few hundred qubits is the practical ceiling. To go further, engineers combine several approaches:

• Lower conductivity cables. Thinner coax and superconducting flexible cables carrying many lines on one substrate cut passive load per line.
• Bigger refrigerators. More pulse tubes, larger plates and higher helium-3 flow add cooling at every stage.
• Multiplexing. Frequency multiplexed readout puts several qubits on one output line; shared control lines cut drive wiring.
• Cryogenic control electronics. Hornibrook and colleagues (Physical Review Applied, 2015) demonstrated a cryogenic FPGA and a switching matrix at 20 mK controlling a semiconductor qubit, explicitly to reduce the lines crossing stages. The catch is that electronics inside the cryostat dissipate their own power.
• Modular machines. Several refrigerators linked by microwave or optical connections, so no single cryostat carries everything.

Error correction makes this sharper, because each logical qubit needs many physical ones; our explainer on quantum error correction, surface codes and thresholds covers that overhead.

From parts to a cold, running system: the step table

The general sequence is below. It describes no single manufacturer's procedure; follow your system's manual, especially for gas handling and pumping, which are system specific and safety relevant. Durations are illustrative.

Because cooldown and warmup each take days, groups pack as much as possible into each cooldown and test components in smaller, faster cryostats first.

How the approaches compare

The cooling hardware depends almost entirely on what the qubit physically is. This compares physical approaches, not companies or products.

This is why "quantum computers are colder than outer space" is only true for some machines. NASA's WMAP mission measured the cosmic microwave background at about 2.725 K. A superconducting processor at 10 to 20 mK is over 100 times colder; a neutral atom vacuum cell sits at room temperature.

A small calculator you can run

The physics reduces to a few formulas. This dependency free Python script reproduces worked examples 1 and 5 so you can change the frequency, per line load or line count. It is a teaching aid, not a design tool.

# cryo_budget.py: thermal photons and a simple wiring heat budget
import math

H = 6.62607015e-34   # Planck constant, J s (exact, SI 2019)
K = 1.380649e-23     # Boltzmann constant, J/K (exact, SI 2019)

def photons(freq_hz, temp_k):
    """Mean thermal photon number in one mode (Bose Einstein)."""
    return 1.0 / math.expm1(H * freq_hz / (K * temp_k))

for t in (300, 4, 1, 0.1, 0.05, 0.02, 0.01):
    print(f"5 GHz at {t:>6} K: n = {photons(5e9, t):.3g}")

# Illustrative 4 K budget using per line loads from Krinner et al. (2019)
COOLING_4K_W = 1.5          # W at 4.2 K, from the paper's Table 1
DRIVE_W, FLUX_W = 1.0e-3, 1.2e-3
HEMT_W = 5e-3               # illustrative HEMT dissipation
QUBITS_PER_READOUT = 8      # illustrative multiplexing factor

for q in (50, 200, 500, 1000):
    hemts = math.ceil(q / QUBITS_PER_READOUT)
    load = q * (DRIVE_W + FLUX_W) + hemts * HEMT_W
    print(f"{q:>5} qubits: {load*1e3:7.0f} mW at 4 K, "
          f"{100*load/COOLING_4K_W:5.0f} percent of cooling")

Run it with any Python 3:

python3 cryo_budget.py

Try setting QUBITS_PER_READOUT to 16, or halving the per line loads to model thinner cable, and watch the ceiling move.

Common mistakes

Confusing 4 K with millikelvin. "Liquid helium temperature" means about 4 K, where a 5 GHz mode holds about 16 thermal photons. The mixing chamber is two to three orders of magnitude colder, and the physics is completely different.

Treating base temperature as operating temperature. An empty refrigerator might reach 6 mK. With wiring and running gates, the mixing chamber settles where cooling matches load; the ETH Zurich group saw 6.1 mK become 8.4 mK from passive wiring alone. Specify cooling power at your operating temperature.

Ignoring active heat load. Signal power in attenuators, DC bias currents and amplifier dissipation depend on the experiment and can exceed passive conduction, especially at the mixing chamber.

Ignoring vibration. Pulse tube compressors shake the system. Rigid compressor lines, poorly decoupled frames or loose cables can heat the mixing chamber and add noise, sometimes at frequencies nobody thought to check.

Ignoring magnetic shielding. Trapped flux raises loss and shifts frequencies. Nearby magnets and steel can matter, and superconducting shields only help if they cool through their transition in a low field.

Forgetting infrared leaks. A gap in light tight shielding or an unfiltered cable lets warm radiation reach the chip. The symptom is shorter coherence with a base temperature that looks fine.

Assuming more attenuation is always better. It lowers thermal photons but raises dissipation at cold stages. Krinner and colleagues explicitly traded noise photon number against mixing chamber heat when choosing their layout.

Where software fits: the control stack is a dependency tree

Safeguard does not do physics, cryogenic engineering or hardware analysis, and its models are single purpose security models, not science tools. The honest connection is narrower. Every dilution refrigerator in a lab is driven by software: gas handling and temperature controller software, instrument drivers, pulse sequencing and calibration frameworks, and data pipelines, usually Python built from dozens to hundreds of open source packages. Those packages carry known vulnerabilities and unpinned versions like any other codebase, and the machines running them often sit on lab networks with remote access. Treating the control stack as a software supply chain, with a lockfile, an SBOM and a scan in CI, is ordinary hygiene, not a quantum specific product.

# Pin the control stack so every lab machine runs the same versions
python3 -m venv .venv && . .venv/bin/activate
pip install pip-tools
pip-compile --generate-hashes requirements.in -o requirements.txt
pip install --require-hashes -r requirements.txt

# Scan the repository for dependencies with known vulnerabilities
safeguard scan --dir .

With --require-hashes, pip refuses any package whose hash does not match the lockfile, so a rebuilt lab computer installs exactly what was tested. The Safeguard scan writes JSON results under ./output, listing components, versions and known vulnerabilities, with reachability analysis ranking what the code actually calls.

A refrigerator does not care which plotting library runs on the control PC, but the person rebuilding that PC after an incident does, because a rebuild pulls every dependency again. The vulnerable packages in a lab's Python control stack are fixed the same way as anywhere else, which is what this section covers.

Fix it without waiting for upstream

Whatever the vulnerability, most of the risk sits in two gaps. The first is the time between a fix being published and that fix actually running in your production builds. The second, and worse, is the time before any fix exists at all, while a maintainer works out a patch and a release schedule. Waiting through both gaps is a habit, not a requirement. Safeguard gives teams three ways to close them.

1. Autonomous remediation: pull requests, not tickets

When a fixed version exists, Safeguard's Griffin models open a tested, policy gated pull request that moves the affected dependency to the safe version. You can run it on one repository or in bulk across every project that carries the vulnerable version. Remediation categories include dependencies, licenses, security posture and findings, and each one is set separately. Reachability analysis ranks the vulnerable code paths your application actually calls first, so the first pull requests you merge are the ones that matter most. Autonomous remediation ships switched off, with every category set to Manual, so your team decides how much runs on its own and what still needs a human approval.

2. Gold Open Source: the safe version in seconds, free

gold.safeguard.sh is free and needs no login. Search any CVE or package to see affected and fixed versions, CVSS, EPSS, CISA KEV status and a recommended safe version, or use the "Am I affected?" checker. The same data is available through a free read API, README status badges and RSS threat feeds.

3. Gold Registry: start from components that are already fixed

Gold Registry provides hardened, drop in replacements for upstream packages and container images. Every Gold artifact is held continuously to zero exploitable critical or high CVEs and a completed license check, with signed build provenance (signed with Safeguard's key and verifiable with cosign, SLSA Build Level 3 as the design target) and a public evidence page you can hand to an auditor. An artifact loses its Gold badge if it regresses. Teams can point their own package directory at the registry so hardened equivalents resolve without changing how developers install anything.

When upstream has no fix yet, or will not ship one, Safeguard first works to get the fix accepted upstream. If the issue is severe or upstream will not move, Safeguard builds a fork, fixes the vulnerability there and publishes a hardened Gold build, so you are not stuck on a maintainer's release schedule. You can also request a hardened, compatibility tested build of a specific package. Ask which packages it covers before you plan around it.

One honest limit: when the vulnerable code lives in a vendor appliance, a hosted service or an operating system runtime you do not build yourself, the fix still has to come from that vendor or distribution. The steps above apply to the open source dependencies and container images you build and ship, which is where most teams carry the largest and least tracked share of their exposure.

See how Safeguard would help

These demos show dependency scanning and remediation on ordinary code. They do not touch cryostats, thermometry or microwave hardware. You can download Safeguard and follow the CLI docs to try them on your own projects.

This is a real CLI session: install, open the interactive terminal and ask it to fix a lodash CVE. The assistant reads the finding and advisory, asks before any command that changes something, edits the dependency, runs install and tests with approval, and pushes nothing. In a lab control repository the same flow updates a vulnerable pinned Python package, and you rerun your calibration tests before merging.

Click through it at your own pace, or open it full screen.

This demo sets up the Safeguard runner on your own machine, approves it with its device code and scans from the terminal. A lab that keeps control code on an internal network can run the scan on a machine inside that network.

Click through it at your own pace, or open it full screen.

This demo connects Claude to the Safeguard MCP server with OAuth and chooses which tools Claude can use. A researcher can then ask in plain language which of the lab's projects carry a given package version.

Click through it at your own pace, or open it full screen.

FAQ

How cold is a quantum computer?

A superconducting processor runs at about 10 to 20 mK at the mixing chamber of a dilution refrigerator; an empty system can reach about 6 mK. Trapped ion and neutral atom machines keep their qubits in vacuum and laser cool the atoms instead.

Why are quantum computers kept so cold?

For superconducting qubits, because thermal photons at their gigahertz frequencies would randomly excite them unless the temperature is well below about 0.24 K, and because the superconductor must be far below its transition to keep quasiparticles rare. At 20 mK a 5 GHz mode holds about six millionths of a photon.

Is a quantum computer colder than outer space?

A superconducting one is. The cosmic microwave background is about 2.7 K, and a mixing chamber at 10 to 20 mK is over 100 times colder. Neutral atom and photonic qubits do not sit in a refrigerator at all.

What is a dilution refrigerator?

A cryostat that cools by moving helium-3 atoms from a concentrated helium-3 liquid into a dilute solution of helium-3 in superfluid helium-4. The crossing absorbs heat, and because helium-3 stays a few percent soluble even near absolute zero, the cooling continues down to a few millikelvin.

Why is helium-3 so expensive and hard to get?

Nearly all of it comes from tritium decay in nuclear weapons programs, at a rate fixed by tritium's 12.3 year half life. CRS reported US supply of about 8,000 liters a year against demand of about 80,000 liters in 2008. Published prices are historical; ask the DOE Isotope Program or a supplier for current terms.

How long does it take to cool down a quantum computer?

Pre-cooling a large dry dilution refrigerator to about 3 K usually takes around a day, and condensing the mixture and reaching base temperature takes hours more. Full cooldown and warmup cycles are counted in days and vary by system and cold mass.

Why do quantum computers have so many cables?

Each superconducting qubit typically needs a drive line and often a flux line, and groups share readout lines. Each line has attenuators, filters and heat sinking at every stage. That wiring forms the familiar "chandelier", and its heat load is a main limit on qubits per refrigerator.

Can quantum computers run at higher temperatures in the future?

Silicon spin qubits have been run at about 1.5 K in the lab, where cooling power is much larger, and ion and atom machines already avoid millikelvin refrigeration for their qubits. Superconducting qubits will stay at millikelvin for the foreseeable future because of their microwave frequencies.

Quick check

Sources

• BIPM: the kelvin, SI base unit definition
• NIST CODATA: Boltzmann constant
• NIST CODATA: Planck constant
• NIST: Ion Storage Group
• NASA: WMAP mission overview (cosmic microwave background temperature)
• Congressional Research Service R41419: The Helium-3 Shortage: Supply, Demand, and Options for Congress (EveryCRSReport copy)
• US DOE Isotope Program, National Isotope Development Center
• Krinner et al., Engineering cryogenic setups for 100-qubit scale superconducting circuit systems, EPJ Quantum Technology 2019 (arXiv:1806.07862)
• Kim et al., A continuously-cooled 3He/4He phase separation refrigerator, 2025 (arXiv:2503.10754)
• Uhlig, Dry dilution refrigerator with 4He-1K-loop (arXiv:1412.3597)
• Schmoranzer et al., Cryogenic broadband vibration measurement on a cryogen-free dilution refrigerator (arXiv:1811.06801)
• Brandl et al., Cryogenic setup for trapped ion quantum computing, 2016 (arXiv:1607.04980)
• Yang et al., Silicon quantum processor unit cell operation above one Kelvin, Nature 2020 (arXiv:1902.09126)
• Bluvstein et al., Logical quantum processor based on reconfigurable atom arrays, Nature 2023 (arXiv:2312.03982)
• Hornibrook et al., Cryogenic control architecture for large-scale quantum computing, Physical Review Applied 2015 (arXiv:1409.2202)

Last reviewed: 25 October 2026.