CMUのベンチマーク、自律的ブラウザエクスプロイトでClaude MythosがGPT-5.5を上回ると判明
CMUのExploitBenchベンチマークにおいて、Claude Mythosプレビューは実世界のV8脆弱性を攻撃する際に人間レベルの自律性を達成。コストは3.6万ドル対3千ドルだったにもかかわらず、OpenAIのGPT-5.5を大きく上回った。
carnegie mellon university's exploitbench evaluates agents on a five-tier scale up to arbitrary code execution within google's v8 javascript engine. claude mythos preview performed comparably to a competent human security researcher, successfully exploiting vulnerabilities like cve-2024-0519 that humans failed to crack for a year. in fully autonomous mode, mythos maintained high performance with a 9.55 score, whereas gpt-5.5 dropped to 4.30. despite the capability gap, the economic barrier remains steep, with mythos requiring twelve times the compute cost of its competitor. anthropic and openai provided api credits for the independent study.