Google, Anthropic and OpenAI Launch Cyber AI Models With Restricted Access

In the first week of September 2026, the three largest Western AI labs made the same move: Google, Anthropic and OpenAI unveiled models with advanced cybersecurity capabilities — and instead of opening them up in the API, put every one of them behind controlled-access programs. It is the first time the industry has treated “hacking capability” as something that requires credentials, not just a credit card.

Quick answer: what was announced?

Google launched Gemini 3.8 Flash Cyber, available through the Fairwind Program for governments, healthcare and telecommunications. Anthropic released Claude Mythos 5.1 only through trusted-access programs. OpenAI confirmed that GPT-6 Astra reached the “Critical” cybersecurity capability threshold and is offering it to testers through the Daybreak Blue program. The common thread: access limited to defenders.

Google: Gemini 3.8 Flash Cyber and the Fairwind Program

Released on September 2 alongside the standard Gemini 3.8 Flash, Gemini 3.8 Flash Cyber was trained, according to Google, with vulnerability fixing prioritized over offensive capabilities. Access runs through the Fairwind Program, aimed at “high-priority defenders” such as governments, healthcare providers and telecom operators. More than 650 partners are already on board, including CrowdStrike, Datadog, Palo Alto Networks and Snowflake. Pricing has not been published.

Anthropic: Mythos 5.1 only for trusted access

Anthropic split its new generation in two. Claude Fable 5.1 may identify vulnerabilities, but penetration testing, exploit generation and binary-based vulnerability scanning are redirected to the Opus models inside controlled programs. Claude Mythos 5.1 lives entirely within those programs, under the Enterprise Frontier Safeguards package: zero data retention combined with state-of-the-art misuse detection.

OpenAI: Astra crosses the “Critical” line

GPT-6 Astra is the first OpenAI system to meet the “Critical” cybersecurity capability threshold under its Preparedness Framework. In evaluations, the model scored 100% on ExploitBench, declined 91.5% of jailbreaking requests, and discovered two zero-day vulnerabilities during testing. That is why the full capability reaches a group of testers through the Daybreak Blue program before any broad release.

Who can use what

Why this matters to you

For businesses, the message is twofold. First: the most advanced AI defense tools will reach official program members first — it is worth watching partners such as CrowdStrike and Palo Alto Networks, which already distribute these capabilities. Second: the same technology that finds flaws in seconds is being restricted precisely because it can exploit them. If you run infrastructure, now is the time to review patching, authentication and monitoring, before those capabilities leak to the other side.

Frequently asked questions

What is Gemini 3.8 Flash Cyber?

It is the cybersecurity-focused version of Gemini 3.8 Flash, released by Google on September 2, 2026, and available only through the Fairwind Program for governments and high-priority defenders.

Can any company use Claude Mythos 5.1?

No. Mythos 5.1 is offered only through Anthropic’s trusted-access programs for cybersecurity and life-sciences work.

What does GPT-6 Astra’s “Critical” level mean?

It is the highest classification in OpenAI’s Preparedness Framework for cybersecurity capability. That is why the model is being released first to testers in the Daybreak Blue program.

At DigitalRadar, we follow security and AI closely. Stay on the radar so you do not miss the next update.