Google’s Gemini 4 Argon launch post contains one line security teams will keep rereading: “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.” That build went to Fairwind Program defenders first. Since October 13, 2026, Argon is also generally available: security teams with a paid Gemini API key can call gemini-4-argon directly, while Fairwind remains Google’s cyber-defense partner program. On Google’s real-world vulnerability discovery eval, Argon scores 85.8%, up from 71.0% for Gemini 3.8 Flash Cyber.
This post covers the cyber numbers, the two access routes, Google’s safeguards, and what the capability means for teams that run APIs. For the model overview, read what Gemini 4 Argon is; for the full rollout timeline, see the Argon release date guide. The practical half is about your own endpoints, and you can do all of it in Apidog.
Gemini 4 Argon cybersecurity numbers
Google published four cyber results for Argon on its Gemini cyber page. These are Google-reported: the discovery and Wiz results come from internal evals, CWE-bench v1 from the public leaderboard and the Gray Swan results from Gray Swan. Harnesses differ, so read the gaps as directional.
| Benchmark | Gemini 4 Argon | Comparison | What it tests |
|---|---|---|---|
| Real-world vulnerability discovery | 85.8% | Gemini 3.8 Flash Cyber: 71.0% | Google internal eval, with source access |
| Wiz Penetration Test Benchmark | 70.9% | Gemini 3.8 Flash Cyber: 58.2% | Black-box pen testing of a running site, no source code |
| CWE-bench v1 (pass@1) | 68% | Three-way tie with Grok 4.7 and GPT-6 Astra; Claude Opus 5.5 at 67% | Leaderboard on Google’s cyber page |
| Gray Swan IPI attack success (k=15) | 0.7% | Lowest on the chart; Kimi K3 highest at 52.7% | Indirect prompt injection; lower is better |
Two details from Google’s evals methodology matter here. The real-world discovery eval ran on “an internal version of the Antigravity harness that is not cyber specialized,” so the 85.8% didn’t come from a purpose-built security agent. And the CWE-bench result is a tie, not a win: the launch post says Argon “ties for first place.” The full Argon benchmark breakdown covers the other 18 rows.
What Argon does for defenders
Google’s summary is direct: “Argon can autonomously find, validate, and patch critical software vulnerabilities.” Each verb maps to a piece of evidence.
- Find. The internal vulnerability benchmark covers exposures across codebases “spanning 20 programming languages.”
- Validate. On Wiz’s black-box benchmark, Argon beat 3.8 Flash Cyber on attack-surface discovery, vulnerability identification, and proof-of-concept evidence, with access only to the running website and its public behavior.
- Patch. The cyber page lists “Automated patching: Generates validated, high-quality code fixes automatically after discovering vulnerabilities.”
Fairwind partners with Argon access can run it inside CodeMender, Google’s specialized code security agent, instead of building their own harness. Earlier cyber models took the same path: Gemini 3.8 Flash Cyber launched with CodeMender in September, and our GPT-5.6-Cyber vs Gemini 3.5 Flash Cyber comparison covers the generation before.
The headline field result came through Wiz’s Scan for Good program. Google says Argon found “a critical vulnerability exposing sensitive personal information across healthcare software used by hospitals worldwide,” one that previous frontier models missed.
Two routes: the paid API and Fairwind
As of October 15, 2026, a security team can reach Argon two ways:
| Paid Gemini API | Fairwind Program | |
|---|---|---|
| Who gets it | Developers and enterprises on a paid tier | Vetted governments, critical infrastructure operators and core technology platforms |
| How you use it | gemini-4-argon (stable) in your own harness |
Standalone or inside CodeMender |
| Price | $2/$10 per 1M tokens through November 30, 2026, then $4/$20; no free tier | Not published |
| Data handling | Paid-tier content isn’t used to improve Google’s products | Zero data retention as a managed model on Gemini Enterprise |
On the API route, Argon supports function calling, structured outputs, code execution and computer use (Preview), with a 1,048,576-token context window shared by input, thinking and output. Rate limits are spend-based too: a Tier 1 project can spend $10 per rolling 10-minute window, so pace fan-out jobs. Our Gemini 4 Argon API guide covers request shapes.
The guardrail-free build is the Fairwind one, and it stays inside a fenced program. Partners accept phishing-resistant MFA, access limited to internal security, incident response, and pen-test teams, per-employee tracking, and no resale. Permitted work covers “authorized threat simulation, reverse engineering, and malware analysis for defensive and academic research purposes”; creating malware isn’t allowed. If you don’t qualify, the Fairwind page points you to CodeMender with publicly available models and Google’s other AI threat defense products.
The safeguards Google described for the wider release
Before broad availability, Google said it was strengthening four safeguard areas:
- Misuse. Argon refuses cyber and CBRN harm “while preserving legitimate, dual-use scientific research, as per our Frontier Safety Framework.” Google monitors internal activations to spot misuse and uses internal and external red teams.
- Prompt injection. Google calls Argon “our most resilient model yet against indirect prompt injections,” backed by the Gray Swan result above.
- Misalignment. Google monitors chain-of-thought and actions and stops execution when necessary. It also urges the industry “to preserve reasoning transparency.”
- Hardened sandboxes. Environments are isolated and sealed before high-risk training and evaluations.
Google’s reasoning for the split: “Safely releasing frontier capabilities at this level requires a phased approach.” It also took part in the U.S. government’s voluntary pre-release access process. General availability came 13 days after the announcement.
What this means for the APIs you run
Here’s our read, and it’s an opinion, not something Google said: frontier discovery ability doesn’t stay inside one program. The guardrail-free build stays with Fairwind partners, but Argon itself is now a paid API call, and other labs keep shipping. When a model can probe “the running website and its public behavior” and find what earlier models missed, your public API surface is the thing being read.
You can’t patch what you haven’t found yet, but you can shrink what an outside prober sees. Start with the boring work:
- Inventory every endpoint from your OpenAPI spec. Import the spec into Apidog so the documented contract becomes your list of what you meant to expose.
- Find shadow endpoints. Diff your gateway’s route table or access-log paths against the spec. Any route that answers but isn’t documented is a forgotten feature or a gap in your docs.
- Require auth on every route. For each endpoint, send a request with no token, an expired token, and another user’s token. Assert 401 or 403 every time.
- Validate input. Send oversized bodies, wrong types, and unexpected fields. Assert 400 or 422, and make sure error bodies don’t leak stack traces.
- Enforce rate limits. Loop a request in a test scenario and assert that you eventually get 429.
- Run it on a schedule. A new route that ships without auth should fail a test the same day, not show up in someone else’s report.
Here’s what those negative tests send and what a healthy API returns:
| Check | What you send | Expected response |
|---|---|---|
| No credentials | Request with the auth header removed | 401 |
| Expired token | A revoked or expired token | 401 |
| Wrong user | User B’s token on user A’s resource | 403 or 404 |
| Oversized body | A payload above your documented limit | 413 or 400 |
| Wrong types | A string where the spec says integer | 400 or 422 |
| Burst traffic | The same request in a loop | 429, ideally with Retry-After |
A post-processor assertion for steps 3 and 4 can be this short:
pm.test("Rejects missing or invalid credentials", () => {
pm.expect(pm.response.code).to.be.oneOf([401, 403]);
});
pm.test("Error body leaks no stack trace", () => {
pm.expect(pm.response.text()).to.not.match(/Traceback|Exception:|\.js:\d+:\d+/);
});
Group these checks into a test scenario per service, run it with Apidog’s scheduled tasks or from CI with the Apidog CLI, and route failures to a notification. Store the expired and other-user tokens as environment variables, not in the request. Apidog is an API design and testing platform, not a vulnerability scanner: it won’t find a memory-safety bug in your parser. It checks, on every run, that the running API still honors the contract you designed, and that’s the surface an outside prober meets first.
If you put a model behind your API
Argon’s 0.7% Gray Swan result is the lowest attack success on Google’s chart, not zero. If your API passes user text, retrieved documents, or tool output to any model, treat the model’s output as untrusted input. Validate tool-call arguments against a schema, scope the credentials a model can trigger, and log every action it takes.
Google’s own safeguard list offers a pattern: watch actions, and stop execution when something looks wrong. At the API layer, that means allowlisting the endpoints a model may call, requiring human approval for destructive operations, and rate-limiting model-triggered calls per key.
Our guide to prompt injection for API teams shows how to test for it, and the Claude Opus 5.5 prompt injection analysis reads a vendor’s injection-resistance number as an API security metric. The same discipline applies to Argon on the public API, where function calling and computer use mean every tool call it makes is a request your API should validate.
How this compares with OpenAI’s Astra
OpenAI’s GPT-6 Astra story was about a threshold: Astra crossed OpenAI’s Critical cyber line, which we unpacked in what Astra’s Critical cyber threshold means for your APIs. Google’s Argon story is about access: a guardrail-free build for vetted defenders first, then general availability on the paid API. On CWE-bench v1, Argon and Astra tie at 68%. Either way, the defensive work above doesn’t depend on which lab ships first.
FAQ
Does Gemini 4 Argon ship without safety guardrails? Only for a narrow group. Google says it released Argon “without cyber guardrails” for trusted defenders and its internal teams. For the wider release, which reached the paid API on October 13, 2026, Google described misuse, prompt-injection, misalignment, and sandbox safeguards.
Do I need Fairwind to use Gemini 4 Argon? No. gemini-4-argon is GA on the paid Gemini API tier, and Google AI Pro and Ultra subscribers get it in the Gemini app. Fairwind is the route for the guardrail-free defender build.
Can my security team use Argon on the paid API? Yes, with a billing-enabled Gemini API key; there’s no free tier. Fairwind’s terms are the ones that explicitly permit threat simulation, reverse engineering, and malware analysis for defensive research.
What is CodeMender? Google’s specialized code security agent. Fairwind partners can run Argon inside it instead of building their own harness.
How does Argon score on CWE-bench v1? 68% at pass@1, a three-way tie with Grok 4.7 and GPT-6 Astra. Claude Opus 5.5 scored 67%.
Is Apidog a vulnerability scanner? No. Apidog designs, mocks, and tests APIs. Use it to keep auth, validation, and rate-limit checks running against every endpoint in your spec.
Your next step
Pick one service, import its OpenAPI spec, and build a scenario that sends every endpoint a request without credentials. If anything that should be private returns 200, you’ve found your first fix. Download Apidog to set up the scenario and schedule it.



