Google's Gemini 4 Argon Hunts Vulnerabilities —and Defenders Get It First
Google’s Gemini 4 Argon Hunts Vulnerabilities —and Defenders Get It First
Section titled “Google’s Gemini 4 Argon Hunts Vulnerabilities —and Defenders Get It First”Google introduced Gemini 4 Argon on September 30, 2026 (Google). It is the company’s first new flagship generation since Gemini 3 shipped in November 2025 (Ars Technica). You cannot use it yet. The first users are a small group of trusted cyber defenders, and they get a version with no cyber guardrails (The Next Web).
Google spent the summer promising a flagship model that never shipped. Gemini 3.5 Pro, previewed at I/O in June, did not arrive (Ars Technica). The company shipped smaller Flash models instead. The gap between Google, Anthropic,and OpenAI grew (Trending Topics). Argon is the reset. It is a frontier model built for defensive security, rolled out first where trust is highest.
The rollout is deliberately narrow
Section titled “The rollout is deliberately narrow”Argon rolls out in phases. Today it reaches trusted testers via Google’s Fairwind Program,a vetted set of cyber defenders, government agencies,and Google Cloud customers (AlphaCorp). Google’s own internal teams use it too (AlphaCorp). Paid API customers and Google AI Ultra subscribers are next in line,with no date announced (The Rundown).
The notable part is what defenders get. Google releases Argon to them without its cyber guardrails,so they can use the full defensive capability (The Next Web). Google says the model can locate software vulnerabilities,validate them,and patch them without human help (Help Net Security).
Wiz already runs Argon through Scan for Good,its free program that finds and fixes high-risk exposures in critical public infrastructure (Google). In an early run,the model surfaced a critical vulnerability exposing sensitive personal information in healthcare software used by hospitals worldwide (Google). Google says previous frontier models missed it (Help Net Security). Google did not name the software or say whether it has been patched (Help Net Security).
The public version will keep guardrails. Google says it plans to scale safeguards in four key areas,including misuse and prompt injection,before launching publicly (CNBC). Google is also participating in the U.S. government’s voluntary pre-release access process for frontier models (AlphaCorp).
The numbers Google published
Section titled “The numbers Google published”Google published a full benchmark set with the launch. These results come from internal tests. No independent party has assessed them yet (Superpower Daily).
| Benchmark | Argon | Comparison | Source |
|---|---|---|---|
| DeepSWE v1.1 (long software engineering tasks) | 77.9% | New high,ahead of GPT-6 Astra, Claude Fable 5.1,and Opus 5.5 | Ars Technica |
| CWE-bench v1 (vulnerability remediation) | 68% | Ties for first place | |
| Internal vulnerability discovery (20 languages) | 85.8% | 3.8 Flash Cyber: 71.0% | AlphaCorp |
| Wiz black-box pentest (live web,no source code) | 70.9% | 3.8 Flash Cyber: 58.2% | AlphaCorp |
| AutomationBench | 51.3% | n/a | Superpower Daily |
| Vals Index (finance,legal,professional) | Leads the field | Ahead of GPT-6 Astra and Claude Fable 5.1 | CNBC |
Two details matter to teams that run security tooling. Argon’s internal vulnerability discovery score covers complex codebases spanning 20 programming languages (Google). Its Wiz black-box score came from probing live web systems with no source code (AlphaCorp). That is how an external attacker sees a system,not how an IDE user does.
Argon also beats OpenAI’s GPT-6 Astra on several coding and knowledge-work benchmarks,per Axios (The Next Web).
The economics target long agent runs. Argon supports up to one million output tokens per run, up from 64,000 (Superpower Daily). That is production tokens,not input context (Superpower Daily). Introductory API pricing is $2 per million input tokens and $10 per million output tokens (The Rundown). Cached inputs cost 95 percent less than standard rates. The price moves to $4 and $20 after the intro window (The Rundown).
Argon is already doing the work inside Google
Section titled “Argon is already doing the work inside Google”Google says engineers use Argon in production today. It helped free hundreds of terabytes of memory across Google data centers without buying new hardware (CNBC). Argon agents are migrating C/C++ codebases to Rust across Google,including more than 800,000 lines in the Fuchsia OS Zircon kernel (Ars Technica).
What to do about it
Section titled “What to do about it”-
Treat AI-assisted vulnerability discovery as its own workload class. Argon is trained to find,validate,and patch flaws. Google is already pointing it at critical infrastructure (Google). When access opens,run it against your own repos before you trust the headline scores.
-
Pick the safety profile to the operator. The same model ships without cyber guardrails to defenders,and with scaled safeguards to the public (CNBC). Apply that rule inside your org: least privilege per role. Full capability only where the blast radius is contained.
-
Read the benchmarks as vendor-reported,not verified. Google’s numbers come from internal tests. Superpower Daily notes they are “not independently assessed” (Superpower Daily). Re-run the key sets on your own data once you have access.
-
Price long runs on cached input. One million output tokens and cheap cache reads target long-horizon agent jobs (Superpower Daily). That mirrors the cached-input economics OpenAI pushed with GPT-6.1 Sol at DevDay (our Sol post.
The pattern
Section titled “The pattern”Two frontier-lab moves this week formed the same arc. OpenAI shelved the model it could not control,then shipped always-on agents (our Astra post. Google shipped its frontier model,no guardrails,to defenders it trusts. Both bets say the same thing: scope,not raw capability is the new competition. Google did not name the healthcare software it found vulnerable,so the benchmark story stays open until it does (Help Net Security).