Claude Mythos、英国AISIのサイバー攻撃シミュレーションで初の完全突破
AnthropicのClaude Mythosプレビューが、英国AI安全研究所(AISI)のサイバーシミュレーションをすべて突破した初のモデルとなり、試行の60%で32段階のネットワーク攻撃を完了した。
the uk ai safety institute revised its forecast for ai cyber capabilities, cutting the expected doubling time from eight months to 4.7 months as models like claude mythos preview and gpt-5.5 outpace projections. mythos preview became the first model to clear both aisi cyber ranges, executing a 32-stage corporate network attack in 6 out of 10 attempts and compromising an industrial control system simulation in 3 out of 10 attempts. independent security firm xbow validated these results, noting a 42% reduction in false negatives for vulnerability detection compared to anthropic's opus 4.6, though performance relied heavily on access to live systems rather than just source code. despite the performance leap, xbow highlighted high costs, noting mythos preview runs five times the price of opus models, raising questions about cost-efficiency for widespread security deployment.