Notes AI and security No. 10
Cyber-gated AI models and who can use them
The most capable vulnerability-finding versions of frontier AI models are not on general sale. Anthropic, OpenAI and Google each offer a public model that refuses or reroutes some security work, and a cyber-gated AI model tier, with fewer restrictions, for vetted defenders. Microsoft runs its security model inside its own service. This note sets out what each lab allows and who can apply, from the terms the labs had published as of 30 September 2026. Programmes like these change often, so check the linked pages before you apply.
What a cyber-gated AI model is
The public and gated versions are often the same model with different safeguards. Anthropic says Claude Fable 5 and Claude Mythos 5 share an underlying model, and that Fable carries extra safeguards. OpenAI says it trained GPT-6 Astra so that its refusal boundary can be made more conservative for users flagged as potentially high risk.
The line usually falls between finding a weakness and exploiting it. Public tiers increasingly help with the first. Exploit development, penetration testing and offensive tooling sit behind an application, identity checks and, in some cases, a nationality or jurisdiction test.
Anthropic's Mythos 5.1 and Cyber Verification Program
When Anthropic launched Claude Fable 5 and Mythos 5 on 9 June 2026, Fable 5 handed cyber requests to Claude Opus 4.8. Anthropic said more than 95% of Fable sessions involved no fallback. Mythos 5 went at first only to Project Glasswing partners, and Anthropic requires 30-day retention for all traffic on Mythos-class models.
Claude Fable 5.1, released on 1 September 2026, moved the line. It "can now be used to discover software vulnerabilities" but not to develop exploits. Its safeguards "still redirect several kinds of dual-use cybersecurity tasks" to Opus models: "penetration testing, exploit generation, and binary-based vulnerability scanning". Anthropic says Claude Code users should see around 60% fewer interventions per session from its cyber safeguards.
Mythos 5.1, the less restricted twin, is for vetted cyber defenders and life scientists. Anthropic says: "Currently, it is only available to a set of US organizations", and that it is working with the US government to widen access to domestic and international partners. Anthropic also runs Claude Security, its codebase scanner in public beta for Claude Enterprise customers, on Mythos 5.1. Users get findings and suggested patches, not direct access to the model.
The Cyber Verification Program is the route for everyone else. Opus and Sonnet models block what Anthropic calls high-risk dual use, "such as vulnerability exploitation or offensive security tooling development". The free, application-based programme gives verified professionals reduced safeguards on those models for legitimate defensive work. Applicants must verify their identity, Anthropic aims to decide within two business days, and organisations on Zero Data Retention are not currently eligible. As of 30 September the programme was not available through Amazon Bedrock, and Anthropic's help page said it did not yet cover Opus 5.5 or Sonnet 5.5, with those models and Mythos-class models to be added "soon". On 21 August 2026 Anthropic had said Mythos-class access would follow.
OpenAI's Trusted Access for Cyber and Daybreak
OpenAI's GPT-6 Astra system card, published on 3 September 2026, rates Astra Critical for cybersecurity, the first OpenAI model at that level. Outside its trusted access programme, OpenAI restricts "scaled agentic vulnerability research and chained exploit development".
The programme is Trusted Access for Cyber, "also known as Daybreak access". Organisations apply for their teams, and individuals verify their identity and request access. Individual members must turn on Advanced Account Security to use Daybreak Blue. OpenAI's API changelog of 7 August 2026 describes two tiers for approved defenders: Daybreak Blue, which OpenAI recommends for most defensive work, and Daybreak Red, which gives separately approved access to purpose-trained models such as GPT-5.6 Cyber for authorised exploit validation, penetration testing and red teaming. OpenAI also says it is "taking additional steps to restrict access by high-risk entities and in high-risk jurisdictions", without listing them in the card.
The card shows what the gate is worth. In OpenAI's tests, Daybreak Blue raised Astra's completion of proof-of-concept exploit creation from 2.4% to 92%, and of cyber red-teaming tasks from 7.4% to 76.9%. The GPT-6.1 Sol addendum of 29 September 2026 also rates that cheaper model Critical for cyber and applies the same safeguards and the same phased Daybreak approach.
Google's Flash Cyber models and the Fairwind Program
Google introduced Gemini 3.5 Flash Cyber on 21 July 2026 for "governments and trusted partners" through CodeMender, its tool for fixing vulnerabilities. Google says it found 55 unique confirmed issues in the V8 JavaScript engine, against 47 for mainline Gemini 3.5 Flash and 36 for Claude Opus 4.6.
On 2 September 2026 Google followed with Gemini 3.8 Flash Cyber and the Fairwind Program, "a limited access program for governments and trusted partners". It names three groups: governments and national cyber authorities, critical infrastructure operators, and core technology platforms. Participants must limit access to internal security, incident response or penetration testing staff and use protections such as multi-factor authentication. Google says Fairwind works with over 650 partners globally. It takes applications through a form on the Fairwind page, which says Google runs background checks on organisations that apply.
Microsoft keeps its model inside a service
Microsoft took a different route. MAI-Cyber-1-Flash, which Help Net Security reported on 27 July 2026 and whose Microsoft post is dated 13 August 2026, runs inside MDASH, Microsoft's multi-agent system for finding, validating and fixing vulnerabilities. Microsoft says the model handles up to 90% of tasks, with GPT-5.4 taking the hardest 10%, and that MDASH scores 96% on CyberGym's any-crash measure. That score is for the whole system, not the bare model. The post does not say who can buy MDASH.
Who can get what, as of 30 September 2026
| Lab | Gated tier | Who can apply | Geography stated |
|---|---|---|---|
| Anthropic | Mythos 5.1 | Vetted defenders and life scientists | US organisations only |
| Anthropic | Cyber Verification Program | Verified defensive professionals | None stated |
| OpenAI | Daybreak Blue and Red (Trusted Access for Cyber) | Organisations and verified individuals | High-risk jurisdictions excluded, not listed |
| Gemini 3.8 Flash Cyber via Fairwind | Governments, infrastructure operators, core platforms | None stated, 650+ partners globally | |
| Microsoft | MAI-Cyber-1-Flash in MDASH | Not stated | Not stated |
What this means for UK and European teams
The strongest tier that is plainly closed to UK organisations is Mythos 5.1, which Anthropic limits to US organisations. The other programmes state no UK exclusion, but OpenAI does not publish its list of high-risk jurisdictions.
Washington can also move the line. On 12 June 2026 the US government applied export controls to Fable 5 and Mythos 5, and Anthropic suspended both models for all users. Fable 5 was available globally again from 1 July. Mythos 5 came back only for a set of US organisations. Our note on AI agent attacks on Hugging Face covers what the UK AI Security Institute saw when it tested some of these models, including Mythos 5 with its cyber safeguards off.
Gating does not remove misuse of the public tiers. Anthropic's September 2026 threat report describes actors running "automated exploit foundries" with Claude. Across the report's cases, the models used were Claude Haiku, Sonnet and Opus, not Fable or Mythos. One workflow aimed at network appliances "yielded more than a dozen possible zero day findings in a single month".
For anyone working on binaries, one detail stands out. Fable 5.1 will help find bugs in source code, but Anthropic still sends binary-based vulnerability scanning to an Opus model instead. In our view, teams that analyse shipped binaries, as Violet RE does, should not plan on the public tier of a frontier model doing that work. They should also expect any AI-generated finding to need the checks we set out in what a finding should tell you.
What to watch
As of 30 September no lab had published a date for its next change. Anthropic says Mythos-class access through the Cyber Verification Program will come "in the near future", and OpenAI says Daybreak access will expand in phases. For the models these programmes sit on top of, see our September 2026 releases roundup.
Sources
- Anthropic, "Claude Fable 5 and Claude Mythos 5", 9 June 2026: https://www.anthropic.com/news/claude-fable-5-mythos-5
- Anthropic, "Introducing Claude Fable 5.1 and Claude Mythos 5.1", 1 September 2026: https://www.anthropic.com/claude-fable-and-mythos-5-1
- Anthropic Help Center, "Real-time cyber safeguards on Claude Opus and Sonnet", viewed 30 September 2026: https://support.claude.com/en/articles/14604842-real-time-cyber-safeguards-on-claude-opus-and-sonnet
- Anthropic, "Bringing the cybersecurity capabilities of Claude Mythos 5 to more defenders", 21 August 2026: https://claude.com/blog/bringing-claude-mythos-5-to-more-defenders
- Anthropic, "Redeploying Claude Fable 5", 30 June 2026: https://www.anthropic.com/news/redeploying-fable-5
- Anthropic, "Detecting and countering misuse of AI: September 2026", September 2026: https://www.anthropic.com/threat-intelligence-report-september-2026
- OpenAI, "GPT-6 Astra System Card", 3 September 2026: https://deploymentsafety.openai.com/gpt-6-astra
- OpenAI, "Addendum to GPT-6 Astra System Card: GPT-6.1 Sol", 29 September 2026: https://deploymentsafety.openai.com/gpt-6-1-sol/cybersecurity-capabilities
- OpenAI, "API changelog", viewed 30 September 2026: https://developers.openai.com/api/docs/changelog
- Google DeepMind, "Introducing Gemini 3.5 Flash Cyber", 21 July 2026: https://deepmind.google/blog/introducing-gemini-3-5-flash-cyber/
- Google, "Proactive cyber defense for governments and enterprises", 2 September 2026: https://blog.google/innovation-and-ai/technology/safety-security/fairwind-program/
- Google DeepMind, "Fairwind Program", viewed 30 September 2026: https://deepmind.google/fairwind-program/
- Microsoft AI, "Introducing MAI-Cyber-1-Flash inside MDASH", dated 13 August 2026: https://microsoft.ai/news/introducing-mai-cyber-1-flash-inside-mdash/
- Help Net Security, "Microsoft unveils MAI-Cyber-1-Flash, promises cybersecurity AI at half the cost", 27 July 2026: https://www.helpnetsecurity.com/2026/07/27/microsoft-mai-cyber-1-flash-ai-model/