Anthropic previews Claude Mythos for cyber defense

This is less a product launch than a controlled capability drop: Anthropic is signaling that Mythos-class models are already beyond normal public-release comfort, especially on cyber tasks.

• –Anthropic says Mythos Preview showed large gains over Opus 4.6 on coding, reasoning, search, and cyber benchmarks, including SWE-bench Pro, Terminal-Bench 2.0, and CyberGym.
• –The company is positioning the model as a defensive-security asset for 40+ partners, not a public API model, which is a strong tell that misuse risk is now part of the release decision itself.
• –The system card’s breadth matters: it covers RSP decision-making, cyber evaluations, alignment, model welfare, and user impressions, so this is as much a governance artifact as a benchmark report.
• –For developers, the practical signal is that next-gen Claude models may ship with tighter access controls and stronger safety scaffolding, especially where autonomous code execution and vulnerability discovery are involved.