On 30 September Google announced Gemini 4 Argon and, buried in the announcement, a decision other labs have so far hedged on. In Google's words, quoted by SecurityWeek: "For trusted defenders and our own internal teams at Google, we'll be releasing Argon without cyber guardrails."
What was announced
Argon goes first to the Fairwind Program, Google DeepMind's early-access scheme for vetted defenders. Decrypt puts the partner count above 650. Paid API customers and Google AI Ultra subscribers follow, with no date given. Introductory pricing is $2 per million input tokens and $10 per million output tokens.
The numbers, with the usual caveat
Decrypt reports 77.9 per cent on DeepSWE v1.1, ahead of Claude Opus 5.5 at 74.2 and GPT-6 Astra at 74.1, and a 0.7 per cent attack success rate on Gray Swan's indirect prompt injection test. On Wiz's penetration test benchmark Argon solved 70.9 per cent of web exploit tasks first time. Those are vendor-selected comparisons: Argon leads 12 of the 18 benchmarks in Google's own table, which also means it does not lead six.
Why this matters outside security teams
Google says the model can autonomously find, validate and patch critical vulnerabilities. Separately, Help Net Security reports on research finding that around half of the vulnerabilities likely found by AI lead to remote code execution, against 26 per cent of those found by other means, and that one such flaw was exploited within four days of disclosure. Put the two together: discovery is speeding up on both sides, and the gating question becomes who is trusted to hold the unrestricted version.
That is a vetting question, and vetting is something a customer can ask about. If a vendor in your supply chain says it uses an unrestricted model internally, you are entitled to ask how access is controlled and logged.
- Assume your patch window shrinks. If the four-day exploitation figure is typical, a monthly cycle is too slow for internet-facing systems.
- Ask security vendors which models they use, under what access terms, and who audits that use.
- Treat benchmark tables as marketing until you have run the model on your own code.
- If you are not in Fairwind, plan on the guardrailed version and decide what evidence you would need before relying on it.
To talk through how AI-assisted vulnerability work fits into your remediation process, email sales@halfteck.com.