Anthropic Warns AI Can Develop N-Day Exploits Within Hours

N-day vulnerabilities are publicly disclosed flaws for which patches are available but some devices remain unpatched. Attackers can compare code before and after a patch to reverse-engineer the weakness. Anthropic said weaponization previously often took weeks, giving companies time to roll out updates in stages. If AI compresses that timeline to hours, medical devices, industrial control systems and systems that are difficult to restart would be hit first.

Anthropic released test results on June 8, 2026, showing that Claude Mythos Preview autonomously developed eight code-execution exploits targeting 18 patched Firefox vulnerabilities. It produced the first in under one hour and completed all eight in about 12 hours. It also generated eight complete privilege-escalation exploit chains from 21 Windows kernel vulnerabilities at a cost of $15,700, highlighting the risk of “N-hour” attacks.