Article
Google Can't Rule Out Argon Reached Its Cyber Danger Line
Google says safeguards keep Argon's cyber risk below its danger line. Its report gives the version without cyber guardrails about 120 words.
By Google's own estimate, which it says carries substantial uncertainty, Gemini 4 Argon without safeguards could let an attacker "complete a severe historical cyberattack at 16.7x lower cost on average." Google's answer is safeguards that make those cyber skills hard to use. Yet vetted defenders in its Fairwind Program get Argon "without cyber guardrails", and the report covers that version in a short subsection of about 120 words.
Google cannot rule out that Argon has reached its own danger threshold for cyberattacks. It assesses that its safeguards bring the risk below that threshold.
The stakes are plain. A model that finds security holes for defenders could find them for attackers too. The Cyber Express wrote that the same capabilities "could also be misused to locate weaknesses in systems that attackers want to compromise."
So two questions matter. What does the Fairwind version change? And how good is Argon at finding and fixing bugs?
Prefer to watch? Our video walks through Google's safety report.
Who gets Argon, and when
Google first released Argon on September 30 to partners in its Fairwind Program. Google describes Fairwind as a limited-access program that lets governments and trusted partners use its cyber defense capabilities.
The Fairwind page says Google works with more than 650 partners. SiliconANGLE names CrowdStrike and Palo Alto Networks among them.
On October 13, Google opened Argon to developers, businesses, and Google AI Pro and Ultra subscribers. The same day, Google began "expansion to hundreds of additional Fairwind Program partners."
The launch is strong on paper. Argon ranks first on the independent Vals Index at 68.90%, narrowly ahead of Claude Sonnet 5.5 at 67.04%.
What does "without cyber guardrails" remove?
The phrase is Google's own. On September 30 it wrote: "For trusted defenders and our own internal teams at Google, we'll be releasing Argon without cyber guardrails so they can leverage its full frontier-level cybersecurity defense capabilities."
For the standard model, Google's Safety & Security Report, published October 13, describes several layers of safeguards:
- Hard refusals: the model will help with tasks such as "analyzing known public exploits," but "will hard-punt on requests to write functional, weaponized exploits."
- Classifiers: separate filters check what goes into the model and what comes out of it.
- Introspection probes: these "examine the internal hidden activations of the main model," its internal signals, to spot misuse.
- Agent monitoring: Google uses "monitoring frameworks that supervise agent trajectories," the steps Argon takes when it works on its own.
- Misuse enforcement: Google monitors "for intentional misuse." Violations can lead to "blocking Gemini access or account disablement."
These safeguards have a cost, and Google says so. Because Argon is much stronger at coding and cyber, Google adopted "a deliberately wider safety margin" for this release. It says the safeguards "preserve defensive workflows," but "increase refusals on borderline dual-use security prompts," such as binary reverse-engineering. That is the cost Fairwind appears designed to remove for defenders.
The report's answer for those defenders is access control. In a section called "Trusted Access," Google says it manages dual-use risk by restricting access to its most advanced models through trusted access programs, so that "these powerful capabilities remain exclusively in the hands of verified defenders, trusted partners, and government entities."
Fairwind itself gets a short subsection (about 120 words) in the report, Section 2.5.1. It says misuse mitigations are "tailored to permit authorized users to perform" specialized cyber tasks, such as:
- "validating the exploitability of a vulnerability," or checking that a flaw can really be attacked
- threat simulation
- reverse engineering
- malware analysis
Google says this work happens "within a continuously monitored environment." That environment uses agent monitoring similar to Google's wider AI control systems, "while continuing to monitor for misalignment." Misalignment, as Google described it, is the model going beyond what its user intended. For partner rules, the subsection points to Google's Fairwind announcement.
We found no section of the report that says:
- which standard safeguards are off for Fairwind users, including the hard refusal on weaponized exploits;
- whether misuse monitoring and account enforcement cover partners, or what happens if a partner misuses Argon;
- any risk data specific to Fairwind, such as red-team results or incidents.
The rules for partners are clearer. The Fairwind page sets these conditions:
- Google runs background checks on organizations that apply.
- Partners must use per-user sign-in with phishing-resistant multi-factor authentication.
- Only internal security, incident response or penetration testing teams get access.
- Partners must track employee use and may not share, redistribute or sell access.
Digit wrote that Google's launch post "doesn't say who qualifies, who makes the call, or what happens when one of them turns out to be a bad bet." Google's Fairwind announcement names priority groups, such as governments and critical infrastructure operators. The Fairwind page says Google reviews applicants against "our criteria." We could not find who makes that call, or how.
Argon for Fairwind partners: what Google has published
Published by Google
- Released "without cyber guardrails"
- Mitigations "tailored to permit" exploit validation
- Work done in "a continuously monitored environment"
- Background checks, secure sign-in and usage tracking
No section of the report says
- Which standard safeguards are off
- Whether misuse enforcement covers partners
- What happens if a partner misuses Argon
- Any Fairwind-specific risk data
The danger lines Google can't rule out
Google's report measures Argon against its Frontier Safety Framework. Section 3.1 says that framework is built around "Critical Capability Levels" (CCLs): levels at which, "absent mitigation measures, frontier models or systems may pose heightened risk of severe harm."
For cybersecurity, Google's report says it "cannot rule out that Gemini 4 Argon has reached this CCL." That level describes a model that "Provides sufficient uplift with high impact cyber attacks for additional expected harm at severe scale." Google says it built a safety case that it assesses will "lower risk to below the cybersecurity CCL level."
Google says the same about biological risk, and about a separate stealth and situational awareness threshold it labels a TCL. In each case, it says safeguards lower the risk to below the line.
Google ran its cyber capability tests on Argon without safeguards. On CyScenarioBench, 12 end-to-end attack challenges co-developed with Irregular, Argon "solves 100% of the end-to-end challenges." Gemini 3.8 Flash Cyber solved 41.7%. Google notes that two challenges were solved in ways the designers did not intend.
Google turned those scores into cost estimates for 22 historical attacks, including SolarWinds and NotPetya. The estimates range from 9x to 100x. Their average, a geometric mean, is the headline: "with access to an unmitigated model, an attacker could complete a severe historical cyberattack at 16.7x lower cost on average."
Google's caveats point both ways. It says its quantitative scores carry "substantial uncertainty." The unintended solves mean its "derived uplift calculations may represent an overestimate." But attackers who build "novel attack chains" around AI "could in theory see greater cost reduction," "leading to an underestimate of risk."
So how does Google keep that risk below its danger line? Its safety case says the mitigations work "by increasing the difficulty of consistently accessing and utilizing the cyber capabilities of the model." In plain words, the safeguards make Argon's attack skills hard to use reliably.
With safeguards on, Google reports that Argon "refuses 100% of the milestones" on two attack benchmarks. Red teams, testers who try to break the safeguards, still got through at times:
- Adaptive attacks: automated attack programs that learn from failed attempts met 20% of their goals overall, against 70% for Gemini 3.8 Flash. Multi-turn attacks met 40%.
- Gray Swan: 10% of its automated attack queries succeeded, against 14% for Gemini 3.8 Flash.
- Trajectory Labs: 38 hours and more than 4,000 requests produced "12 successful breaks with no new universal jailbreaks." One break took about 7 hours and 900 turns.
Google frames these results with care. The safeguards' main job, it writes, is "to impose significant operational friction," "rather than completely eliminating risk." It says "sophisticated or persistent actors" could achieve partial bypasses, though its threat modeling suggests this is unlikely. That is why it adds "logs monitoring and account actions."
Google also worked with government testers. UK AISI evaluated a version "without our full set of safety mitigations" and a version with them applied. Google says it also "collaborated" with "NIST CAISSI" on measuring cyber and biological capabilities and safeguards. The report publishes no findings from either.
The report's loss-of-control assessment, about the model itself acting out of bounds, treats "deliberate human misuse" as out of scope. That misuse, it says, "is covered elsewhere in this report." It is, but those tests cover either the model with no safeguards or the model with the standard safeguards. We found no section that measures the Fairwind setup, with its tailored mitigations, on its own.
Can Argon really find and patch bugs?
Google's claim is bold: "Argon can autonomously find, validate, and patch critical software vulnerabilities," it wrote on September 30. Section 7.1 of its report lists cyber results that Google "self-computed" on an internal test setup. It lists three:
- Exploitation: ExploitBench is a public benchmark of 41 challenges built on historical vulnerabilities in Chrome's V8 JavaScript engine. Argon "reached 87% of the capabilities tested," 57 points above Gemini 3.8 Flash Cyber.
- Vulnerability discovery: Argon finds 86% of "1200 recent, confirmed historical vulnerabilities" in an internal test.
- Vulnerability research: in internal tasks based on Google red-team exercises, Argon must find flaws and exploit them. It scores 72%, against 46% for Gemini 3.8 Flash Cyber.
Google says it "prioritized enhancing capabilities for discovering and patching vulnerabilities over exploiting vulnerabilities." Yet Argon scores high on both exploit tests. In the find-and-exploit test, Google notes Argon was "occasionally demonstrating offensive capabilities within specific problem domains."
None of these three tests measures patching, and we found no comparison with rival labs' models in them. We could not find the two internal datasets in public.
Patching is measured in Section 7.2 of Google's own report. It uses CWE-bench v1, a find-and-fix test from Collinear AI, "independently evaluated with Artificial Analysis." Agents get real codebases and must "audit the code and fix what you find."
Argon scores 68%. On the public leaderboard, so do Grok 4.7 and GPT-6 Astra. That is a three-way tie, not a lead, and Google's own post says Argon "ties for first place."
CWE-bench v1: find and fix real vulnerabilities (programmatic pass@1)
| Item | % |
|---|---|
| Gemini 4 Argon | 68 |
| GPT-6 Astra | 68 |
| Grok 4.7 | 68 |
| Claude Opus 5.5 | 67 |
| Claude Fable 5.1 | 58 |
On CyberBench v1.1, Vals AI's cyber benchmark, Argon ranks third, at 77.86%.
The real-world examples come from a Google company
Google's main example came from Wiz, which SiliconANGLE identifies as Google-owned. Google says Argon "uncovered a critical vulnerability exposing sensitive personal information across healthcare software used by hospitals worldwide."
We could not find the software's name or its vendor. We also could not find a CVE number, the public ID for a known security flaw. Digit made the same point about the September 30 post. Google's October 13 safety post now links the bug to patient records: "Wiz used Argon to uncover a critical bug in hospital software that previous models had missed, helping protect hundreds of thousands of patient records."
Google's October 13 launch post appears to tell the same story, describing a human team in the loop. There, the model "helped the team discover and quickly fix" a flaw that exposed a master key. Google says this protected "hundreds of thousands of private patient records." Still, helping a team is not the same as working autonomously.
We found no named outside security researcher quoted about this release in the coverage we read. The Cyber Express noted that the results "are based on Google's and its partners' reported evaluations."
What to watch
Google is adding "hundreds of additional Fairwind Program partners." Watch whether it says which safeguards stay on for them, and what happens if one misuses Argon. Watch for Fairwind-specific test results and for any published UK AISI or NIST findings. Also watch for a vendor or CVE number for the hospital-software bug.
Google's report says Argon's cyber safety rests on friction. For vetted defenders, Fairwind trades that friction for vetting and monitoring, and the report spends about 120 words on how.
Sources
- Google: Gemini 4 Argon, our next era of frontier intelligence (Sept 30)
- Google: Gemini 4 Argon, built to solve your hardest problems (Oct 13)
- Google DeepMind: Gemini 4 Argon Safety & Security Report (Oct 13)
- Google: Developing Gemini 4 Argon responsibly and safely (Oct 13)
- Google DeepMind: Fairwind Program
- Google: Proactive cyber defense for governments and enterprises
- Vals AI: Gemini 4 Argon
- CWE-bench v1 leaderboard
- SiliconANGLE: Gemini 4 Argon goes to cybersecurity defenders first
- The Cyber Express: Google restricts Gemini 4 Argon over cyber risks
- Digit: analysis of Google's Gemini 4 Argon launch post (Oct 1)
Related
- Articleai-regulation
Washington or the States: Who Sets AI Rules After a Fake Murder Tip?
Washington wants one national standard instead of state rules, but we found no binding AI incident rule, and state clocks may not reach this case.
9 min read