Endor Labs benchmarked Claude Fable 5 on 200 real-world vulnerability-fixing tasks for the Agent Security League and found mixed results. Overall, the model performed only moderately well, scoring 59.
Claude Mythos Preview \ red.anthropic.com
news
red.anthropic.com
IrregularChat: AI & Autonomy
Apr 9
On April 7, 2026, Anthropic announced the Claude Mythos Preview, a general-purpose language model excelling in cybersecurity tasks. This launch is complemented by Project Glasswing, aimed at enhancing
Claude Mythos Preview \ red.anthropic.com
news
red.anthropic.com
IrregularChat: Full Stack Dev
Apr 9
On April 7, 2026, Anthropic announced the Claude Mythos Preview, a general-purpose language model excelling in cybersecurity tasks. This launch is complemented by Project Glasswing, aimed at enhancing
Claude Mythos Preview \ red.anthropic.com
news
red.anthropic.com
IrregularChat: Purple Team
Apr 8
On April 7, 2026, Anthropic announced the Claude Mythos Preview, a general-purpose language model excelling in cybersecurity tasks. This launch is complemented by Project Glasswing, aimed at enhancing
CyberGym is a comprehensive framework developed by researchers at UC Berkeley to evaluate AI agents' capabilities in real-world cybersecurity tasks, featuring 1,507 benchmark instances from 188 softwa