Anthropic has announced Claude Mythos Preview, a model so capable at offensive cybersecurity that the company decided not to release it to the public. During red-team testing, Mythos Preview discovered thousands of high-severity zero-day vulnerabilities — previously unknown security holes — across every major operating system and web browser. It also became the first AI model to complete a complex 32-step corporate network attack simulation called 'The Last Ones,' succeeding in 3 out of 10 attempts and completing an average of 22 out of 32 steps across all tries.
Rather than shelving the model, Anthropic launched Project Glasswing: a controlled program that gives early access to Mythos Preview specifically to help patch vulnerabilities, not exploit them. Project Glasswing partners include Amazon Web Services, Apple, Google, JPMorganChase, Microsoft, and Nvidia. The idea is to use the model's offensive capabilities defensively — letting it find holes so engineers can close them before malicious actors do. Anthropic CEO Dario Amodei met with White House officials to brief them on the model's implications.
The announcement has sparked debate in the security and AI communities. Critics argue that even restricted access models create dangerous precedents — if Mythos Preview can find zero-days at scale, so can a future, less safety-conscious lab's model. Supporters counter that proactive vulnerability disclosure backed by AI is exactly what the internet needs, and that Anthropic's transparent approach — publishing evaluation results from the UK's AI Safety Institute — sets a better standard than quietly deploying dangerous capabilities. The Foreign Policy analysis described Mythos as 'changing the cyber calculus.'
For students learning about AI, this story illustrates one of the field's most pressing tensions: capability and safety advancing together. Mythos Preview is not a general-purpose assistant — it is a specialized research tool with a narrowly controlled rollout. Understanding why Anthropic made this call, and what 'responsible deployment' looks like when a model is genuinely dangerous, is essential context for anyone building or studying AI systems. The era of 'release it and see what happens' is giving way to something more deliberate.