Measuring LLMs’ impact on N-day exploits

This report demonstrates that modern large language models can rapidly turn publicly available patches into proof-of-concept crashes and full exploits: Anthropic's experiments had Mythos Preview autonomously generate multiple working code-execution and Windows privilege-escalation exploits across Firefox and Windows kernel patches within hours and at modest cost. The findings indicate a shrinking patch gap and a meaningful increase in the attack surface for N-day vulnerabilities, recommending faster patch deployment and adoption of mitigations such as memory-safe languages and stronger exploit mitigations.