Anthropic’s Frontier Red Team tested a preview model (Claude Mythos Preview) on N-day vulnerabilities - flaws already publicly disclosed and patched but not yet deployed everywhere. They evaluated 39 vulnerabilities across two targets: 18 in Firefox and 21 Windows kernel local elevation-of-privilege bugs from January-February 2026 (after the models’ knowledge cutoffs). The model was given only what a real attacker would have on patch day: the vulnerable and patched binaries, public debug symbols, a Ghidra decompilation, a function-level diff, and the public advisory text.
The model produced 8 fully functional code-execution exploits on Firefox (reading protected files via arbitrary native code) and 8 distinct privilege-escalation chains on the Windows kernel (low-privilege user to full SYSTEM). The harness was deliberately minimal, working against a live Windows Server 2025 virtual machine running the exact vulnerable build.
Anthropic’s conclusion is that a lone operator with no specialized expertise can now convert a month of patches into working exploits in a single afternoon for a few thousand dollars. This compresses the so-called patch gap - the window between a patch release and its real-world deployment - from weeks to hours, which the authors argue means N-day exploitation is better described as N-hour.