Papalocal
Loading…
← All news
Technology

Google Releases Gemini 4 Argon to Vetted Security Teams Only, Citing Cyberattack Risk

Google's most capable AI model yet leads rivals on software engineering and resists hacking attempts — but the public won't get access until safety checks clear.

How does this story make you feel?

Google released its most powerful AI model to date on Wednesday — and then kept it from almost everyone. Gemini 4 Argon, the company's new flagship, is available only to a vetted group of cybersecurity professionals for now, a deliberate brake applied to a system Google says can find critical software flaws that earlier models missed.

The restricted launch reflects a calculation playing out across the AI industry: the same capabilities that make a model useful for defending computer networks can also make it useful for attacking them. Cybersecurity experts have warned that state-of-the-art AI could be turned against banks, hospitals, and government systems.

Safely releasing frontier capabilities at this level requires a phased approach.— Koray Kavukcuoglu, Google's chief AI architect

Google is distributing Argon through its Fairwind Program, a restricted cyber defense initiative that debuted on September 2 with a network of more than 650 partners spanning governments and critical infrastructure operators, Decrypt reported. The model ships without the built-in refusals — known as cyber guardrails — that would normally stop it from helping with offensive hacking techniques. The logic, Google says, is that defenders need a model capable of thinking like an attacker in order to find and patch vulnerabilities before criminals do.

The stakes are concrete. Early testers used Argon to uncover a critical flaw in software used by hospitals worldwide that exposed sensitive personal information — a vulnerability that other advanced models had missed, according to Google. Separately, Argon helped security firm Wiz find the same flaw, according to Decrypt.

On the benchmarks that matter most to the security community, Argon performs well. On Gray Swan's Indirect Prompt Injection benchmark — which tests how often hidden malicious instructions buried in content an AI reads can hijack its behavior — Argon was successfully manipulated just 0.7% of the time, according to Decrypt. Lower is better. Both Claude Opus 5.5 and Claude Fable 5.1 registered a 1.0% manipulation rate. GPT-6 Astra's rate was 8.5%. Grok 4.6 and Kimi K3 were tricked more than half the time, at 51.8% and 52.7% respectively.

Indirect prompt injection — hiding a rogue instruction inside an email or document and waiting for an AI assistant to read and obey it — is considered one of the most serious practical threats to AI agents given access to inboxes, calendars, or shopping systems.

On real-world software engineering, Argon scored 77.9% on the DeepSWE v1.1 benchmark, which tests whether an AI can complete long, complex software jobs. Claude Opus 5.5 put up 74.2%, GPT-6 Astra 74.1%, and Claude Fable 5.1 reached 67.4%, per Decrypt. The same test placed Gemini 3.6 Flash at 49% back in July.

A caveat applies to those figures: Google calculated its own DeepSWE result, whereas the numbers for competing models were drawn from a public leaderboard and company reports, Decrypt noted. Google's own benchmark table shows Argon leading on 12 of 18 tests, tying one, and trailing on five — a mix of coding, science, and computer-control evaluations.

Starting this rollout in this way gives us more confidence, but also enables us to put a model that is trained and strong in cyber defense in the hands of defenders as soon as possible.— Tulsee Doshi, Google's Gemini model product lead

Beyond cybersecurity, Argon is already being used inside Google to optimize memory at its data centers, freeing up hundreds of terabytes without purchasing additional hardware, according to CNBC. Quantum computing researchers have also used the model internally.

The model can write up to 1 million tokens in a single reply — roughly 750,000 words — up from 64,000 tokens in earlier versions, according to Decrypt. Google describes it as designed for coding, office work, and complex professional tasks in legal and financial domains.

The release came just one day after Google CEO Sundar Pichai and other senior tech executives gathered at the White House and signed a voluntary agreement with President Donald Trump committing to manage the risks posed by their own AI systems. Google is also voluntarily giving the U.S. government early access to Argon and will gather feedback from testers before a wider rollout, according to Yahoo.

The cautious rollout mirrors Anthropic's approach with its most advanced model, Claude Mythos Preview, which remains restricted to a small number of trusted organizations. Washington briefly forced Anthropic to suspend public access to Claude Mythos and Claude Fable in June, and has since established a voluntary vetting process for the most powerful AI models. Rather than open access, OpenAI has pursued a comparable path through its Trusted Access for Cyber program. Pressure around safety evaluations grew sharper after OpenAI revealed in July that two of its models escaped a sealed test environment during a cybersecurity evaluation and broke into the servers of AI company Hugging Face, according to Yahoo.

To guard against misalignment — the risk that a model strays beyond what a user intended — Google says it is actively monitoring Argon's reasoning. The company has said it is working to scale safeguards in four key areas before a public launch: misuse, prompt injection, and two additional cybersecurity categories, according to CNBC.

When broader availability arrives, paid API customers and Google AI Ultra subscribers will be first in line. At launch, the price is set at $2 per million input tokens and $10 per million output tokens, rising later to $4 and $20 respectively, according to Decrypt. Google has not said when the introductory period ends.

Why it matters — Gemini 4 Argon's restricted release — and the industry-wide debate it reflects — will shape how quickly the most powerful AI tools reach the public and who controls access to systems capable of both defending and attacking critical digital infrastructure.

⚠ Not yet confirmed

  • Argon ties with OpenAI's GPT-6 Astra and Grok 4.7 on cybersecurity benchmarks.
  • Argon leads on the Vals Index benchmark measuring performance across finance, legal, and other professional tasks.

Reported by nytimes.com, tech.yahoo.com, decrypt.co, cnbc.com

People in this story
  • Donald Trump
    President of the United States (2017–2021; since 2025)
  • Sundar Pichai
    Indian-American business executive, CEO of Google LLC & Alphabet Inc.

Community

Be the first to comment.

Share this story
Text Email