Measuring LLMs' impact on N-day exploits

Executive summary: Anthropic's Frontier Red Team evaluated how advanced large language models accelerate 'N-day' exploit development by turning public patches into working proof-of-concept crashes and full exploits for Firefox (SpiderMonkey) and Windows kernel vulnerabilities. Across their tests, Mythos Preview produced multiple PoCs and eight distinct end-to-end privilege-escalation exploits within hours and at costs of a few thousand dollars, demonstrating that LLMs can dramatically shorten the time and lower the cost to weaponize disclosed patches and thereby increase real-world risk during the patch gap.