By Hamza Faisal

Anthropic’s most restricted AI model is now hunting bugs for companies that could never get it directly. Cognition said on Tuesday that Claude Mythos 5.2 is now available inside Devin Security, its managed vulnerability-scanning service — with more than $100 million in credits behind the offer.

Mythos is the model Anthropic will not sell to just anyone: the same model as the generally available Claude Fable 5.2, but with cyber and biology safeguards lifted, and access limited to organisations vetted through Anthropic’s Cyber or Life Sciences Verification programmes. With Devin Security, Devin runs the scans itself, so a security team gets Mythos-grade scanning without its own Mythos access.

Mythos, without the Mythos waitlist

Devin Security scans a company’s code, ranks flaws by exploitability, and prepares fixes — a workflow Cognition calls an autonomous defence factory. Devin Security 2.0 added deep scans for complex, chained exploits across large codebases at its October 26 launch, powered by Claude Mythos and other frontier cyber models.

What is new this week is the model: Mythos 5.2, released on November 9 alongside Fable 5.2, now powers those scans for eligible enterprise customers, with applications open for companies without access. Fable 5.2 was already in Devin Desktop and Devin CLI from launch day; the security product is where the restricted sibling lands.

The credits are the hook: $100 million-plus in Devin credits to find and fix companies’ most critical vulnerabilities with Mythos 5.2, eligibility criteria and conditions attached. The credits are for its service, not cash — and Cognition has published no per-customer cap, expiry or allocation.

The 73.5 benchmark — and the fine print

Cognition also published the first numbers for the new combination. On its internal cyber benchmark — 25 pinned repositories holding 1,781 publicly reported vulnerabilities, three runs each — Mythos 5.2 scores a mean of 73.5, up from 56.9 for Mythos 5.1 and more than double the 32.7 of Claude Opus 5.5, Anthropic’s flagship general model.

The caveats matter. The benchmark is Cognition’s own, run on its own test set, and nothing about the figures has been independently verified. A score on pinned, publicly known vulnerabilities is not the same as finding a live zero-day in a company’s private codebase — and Cognition’s October announcement carried no benchmark at all for the Mythos deep scans, making this week’s number the first public yardstick.

Even so, the direction is clear: the best vulnerability-hunting models keep improving, and the gap between restricted and general models is widening — Mythos 5.2 more than doubles Opus 5.5 on Cognition’s test.

Why defenders are getting the good model first

The launch fits Anthropic’s deliberate strategy: the model is too capable at discovering software vulnerabilities to release widely, so it rations access to vetted defenders — first through Project Glasswing, its critical-infrastructure coalition, and now through managed services like Devin Security, where the model scans and the customer never touches it directly.

That rationing is the point: every vulnerability Mythos finds in a customer’s code is one an attacker has not found first — assuming the fix lands before the disclosure. Anthropic has acknowledged fewer than 1% of the vulnerabilities Mythos has surfaced have been fully patched so far, in part because it paces submissions to avoid overwhelming maintainers.

For security teams, the calculation is simpler: the most powerful vulnerability-hunting model Anthropic has built is now available as a service, with the first $100 million of scanning effectively free. The question is no longer whether they can get access to Mythos — it is whether their code can survive what it finds.