OpenAI, Google, Anthropic unveil AI cyber models; most stay gated
Astra is the first model OpenAI rates 'Critical' for hacking; Google's cyber model stays gated; Anthropic unblocks defensive vuln-finding on Fable 5.1.
Copy markdown
Astra crosses the 'Critical' line
OpenAI designated Astra its first 'Critical' cyber model: it can find and exploit zero-days across well-defended systems with no human guiding each step. It scored a perfect 100% on ExploitBench, discovered and used two real zero-days during an eval (now being disclosed to maintainers), and refuses 91.5% of malicious-cyber requests versus 59% for GPT-5.6 Sol. Access is alpha-only via the Daybreak Blue program, defensive use first.
Gemini 3.8 Flash Cyber is defense-first — and API-locked
Google's cyber variant tops the CyberGym vulnerability-discovery benchmark and hits 47.2% pass@1 on CWE-Bench auto-patching; Chrome's security team says it produced 2.6x more correct Chrome patches than the best commercial models. It is not in the standard Gemini API — you get it only through the new Fairwind program, gated to trusted governments, critical-infrastructure operators, and software maintainers.
The one you can use today: Fable 5.1 now allows defensive vuln-ID
Anthropic flipped its usage policy: Fable 5.1 — already GA in the API — may now be used to identify software vulnerabilities, the defensive work it previously refused. Penetration testing, exploit generation, and binary-based vuln scanning still get redirected to Opus models. If your security tooling was hitting refusals on legitimate vuln-finding, it is worth re-running.
Enterprise Frontier Safeguards: keep your data, keep the guardrails
Anthropic's new EFS lets enterprises store logs and run misuse review on their own cloud (AWS, Google Cloud, Azure) instead of Anthropic's, pairing zero data retention with cross-session misuse detection. It rolls out in phases starting this fall across Claude Code, Enterprise, and Platform; eligible customers get zero-retention Fable 5.1 until it lands.
Mythos 5.1 sits behind a verification wall
Anthropic's stronger Mythos 5.1 is trusted-access only — the Cyber Verification Program for defensive security professionals and a US-only Life Sciences Verification Program. The top-tier cyber and bio reasoning won't appear on a normal API key; you apply and get vetted first.
Expect more false refusals, even where it's now allowed
All three labs shipped heavier classifiers, chain-of-thought monitoring, and tighter limits on higher-risk accounts. OpenAI openly warns its safeguards 'may erroneously flag legitimate activity as cyber misuse' — so legitimate security-research prompts are likelier to trip a block, even on the paths that were just unlocked.