Anthropic opens Mythos 5.1 to vetted security teams in three tiers
Anthropic said on the 6th of October that approved security teams can now use its strongest models with fewer blocks.
Anthropic expanded its Cyber Verification Program on the 6th of October 2026, according to the company's announcement. Approved security teams can now use Claude Mythos 5.1, Opus 5.5 and Sonnet 5.5 with fewer blocks on cyber work.
Anthropic's public models block most cyber work, the company wrote. The same skills that help defenders fix a flaw can help an attacker use it.
Why it matters
Security teams can now get help from the best Claude models on real attack and defence work. In Anthropic's tests, every such task was blocked at the first step without the program, the post says.
Key facts
The program has three levels, according to the announcement.
Defense Access is for defence work, such as stopping an attack or studying harmful software. Anthropic said it aims to answer applications within a few days.
Red Team Access adds attacks on a system, with the owner's okay, to find weak spots. It is open to groups only, and reviews take a few weeks, the post says.
Specialized Access is for a few groups that test flight systems, power grids and bank networks. Anthropic said it reviews each one together with the US government.
At the Red Team level, Claude Opus 5.5 finished 34 of 50 attack tasks in Anthropic's tests, the company said.
Partners in Anthropic's Glasswing project found at least 129,000 software flaws between April and July 2026, according to the post. Anthropic said the real number is probably at least five times higher.
What's next
Groups apply on Anthropic's website, the company said. Anthropic will check each one and keep their data to watch for misuse.