Anthropic's "Claude Mythos" too risky for public release: only vetted partners get access

San Francisco, May 7. AI lab Anthropic announced Claude Mythos Preview on April 7, 2026 — a frontier model whose hacking capabilities have been deemed too dangerous for general release.

Alarming test results

In pre-release testing, Mythos identified thousands of unknown zero-day vulnerabilities. More alarming: in over 83% of cases, the model produced working exploits on the first attempt. This crosses a threshold AI safety researchers had previously identified as dangerous.

Limited release under "Project Glasswing"

Anthropic released the model under "Project Glasswing" to approximately 50 vetted partners — including Apple, Amazon, JPMorgan Chase, Palo Alto Networks, and several government agencies. Each partner signed strict usage agreements.

A $100 million security fund

Anthropic has committed up to $100 million in usage credits and $4 million in cash donations to open-source security organisations over five years. CEO Dario Amodei said: "We have to take an active role in preventing the misuse of this technology."

OpenAI's response

In May, OpenAI followed with GPT-5.5-Cyber, granting limited preview access to vetted EU cybersecurity teams. The field is fast becoming the new frontier in "defensive AI".

The ethical question

Experts worry that such models could be a "treasure map" for traditional cyber-attackers. A professor at MIT's AI Lab said: "A model that works in 83% of cases is a national security concern. Sharing it with a few companies is part of the solution, but not the whole answer."

Implications for India

India's cybersecurity authority CERT-In has begun discussions with the US government and Anthropic. CERT-In argues that "advanced AI benefits" should be globally accessible, not limited to a handful of large firms.