Google just built the most powerful cybersecurity AI on the market, stripped out the safety restrictions, and handed it to governments and corporate partners — while telling the public to wait indefinitely for access.
Gemini 4 Argon, unveiled Wednesday, leads rival models on 12 of 18 benchmarks Google disclosed and ships what the company calls "without cyber guardrails" — the built-in refusals that normally prevent an AI from assisting with hacking. The logic, as Decrypt reported, is that defenders need a model that can think like an attacker. The catch is that the same skill cuts both ways, and Google admits as much, calling a phased rollout "the only safe path."
Safe for whom? Argon is available right now only through Google's Fairwind Program, which launched September 2 with more than 650 partners including governments and critical infrastructure operators, according to SiliconANGLE. CrowdStrike and Palo Alto Networks are in. The U.S. government is in — Google says it is participating in the government's voluntary pre-release model access process. You are not in. There is no timeline for general availability.
This is the same Google that spent years building censorship infrastructure at the behest of the Biden administration, throttling dissent and suppressing speech it deemed dangerous. Now it wants Americans to trust it as a digital guardian. The company that censored you for questioning regime orthodoxy is building an AI that can write working exploits against real web applications — and its idea of safety is keeping that tool locked behind a velvet rope shared with the state.
Google's own chief AI architect, Koray Kavukcuoglu, wrote that releasing capabilities at this level "requires a phased approach," as SiliconANGLE reported. Separate monitors watch Argon's chain of thought and can stop it if it goes beyond what a user intended. Translation: the model has a kill switch, and someone other than you controls it.
The benchmark claims deserve scrutiny. Google computed its own DeepSWE score of 77.9%, while rivals' numbers came from public leaderboards and company reports, as Decrypt noted. Bloomberg published a less-than-stellar report quoting a Google employee who said "the model struggles to handle certain coding tasks," per Wccftech. VentureBeat counted that Argon leads outright on 12 benchmarks but loses outright on five — hardly the clean sweep the marketing implies.
The real signal is ownership and access. Wiz, the Google-owned cloud security firm, is already using Argon in its free Scan for Good program and claims it found a critical flaw in healthcare software used by hospitals worldwide that earlier models missed, according to SiliconANGLE. No specifics were provided. Google is both the builder and the primary beneficiary of its own security tool, and it is steering the most powerful version toward government and corporate partners first.
Ars Technica framed the launch around the fact that "you aren't allowed to use it yet." That is the story. An American company, subsidized by American infrastructure, employing American engineers, builds a tool with no safety limits — and gives it to the state and the Fortune 500 while the public gets a waitlist. The guardrails come off for the powerful. They stay on for you.
