Anthropic unveils Project Glasswing cybersecurity initiative with Mythos model

Anthropic announced Project Glasswing, a new cybersecurity initiative that includes an unreleased frontier AI model called Mythos Preview. The company says the model achieved 93.9 percent on the SWE-bench Verified benchmark, compared with 80.8 percent for its current flagship Opus 4.6 model. On the more challenging SWE-bench Pro benchmark, Mythos Preview scored 77.8 percent versus 53.4 percent for Opus 4.6. The improved performance suggests significant advances in the model's ability to perform complex software engineering tasks. Anthropic stated it considers the model too dangerous for general release due to its advanced capabilities in areas that could enable malicious use. The company said it is developing the model under controlled conditions with security researchers. The announcement comes as AI labs face increasing scrutiny over the potential misuse of powerful models, particularly in cybersecurity contexts. Anthropic has previously advocated for responsible scaling policies that restrict access to models with certain dangerous capabilities. Project Glasswing represents Anthropic's attempt to develop defensive AI tools while maintaining safeguards against the release of systems that could be exploited for offensive purposes. The company has not provided a timeline for when or if Mythos will be made available commercially.