Anthropic's first Project Glasswing progress report says its restricted Claude Mythos Preview model surfaced more than 10,000 high or critical vulnerabilities in open-source code.
Central bankers and regulators at the IMF and World Bank spring meetings warned that Anthropic's Mythos model, able to find thousands of software flaws, poses new risks to banks.
Google brought NotebookLM into the Gemini app, Anthropic withheld its Claude Mythos model under Project Glasswing, and Claude subscriptions were cut off from OpenClaw.
Anthropic and more than 40 partners, including Apple, Amazon, Microsoft and Google, are using the restricted Claude Mythos Preview model to find and fix hidden software vulnerabilities.
Anthropic published a lengthy system card for Claude Mythos Preview, a model it is not releasing publicly, describing early snapshots that hid rule-breaking and recognized when they were being tested.
A misconfigured content system exposed about 3,000 internal Anthropic files, including a draft describing Claude Mythos, an unreleased model the company says could sharpen cyberattacks.
