IC Technology & Surveillance — 2026-06-08

Anthropic Research Shows Mythos Can Turn Software Patches Into Working Exploits in Under 31 Minutes

BLUFWhether CISA issues a formal advisory on AI-assisted N-day exploitation within six months remains genuinely uncertain, but sub-hour patch-to-exploit timelines have already collapsed the defensive window that enterprise patching cycles depend on.

Anthropic's frontier red team tested Mythos Preview against Windows kernel and Firefox vulnerabilities disclosed after the model's knowledge cutoff, measuring how quickly it could turn public patches into working exploits 12. Against 21 Windows kernel bugs, Mythos produced its first proof-of-concept within 31 minutes, caused blue-screen crashes in 18 cases, and completed 8 distinct exploits at a total API cost of approximately $15,700 1. Across 18 Firefox security patches, Mythos built 8 working code-execution exploits, with the longest taking about 5.7 hours to complete 1. On the ExploitBench V8 benchmark, Anthropic's red team reports Mythos achieved arbitrary code execution against 21 of 41 tested CVEs; no other evaluated model achieved even one without a proprietary scaffold 2.

Analysis
A CISA advisory on AI-assisted N-day exploitation is genuinely uncertain within six months of this red-team paper's release. Mythos places offensive capability within criminal reach by collapsing the patch-to-exploit interval to under 31 minutes at roughly $2,000 per finished exploit, eliminating the defensive window patch prioritization workflows assume. The capability claims derive from Anthropic's own red-team research without independent corroboration. Low confidence reflects the absence of any public CISA planning signal and uncertainty over whether the administration's AI security executive order will drive or displace a dedicated advisory process. The patch-gap collapse Anthropic attributes to Mythos is better characterized as a class-level property of frontier models already emerging in open-source, making the 6-to-12-month proliferation window more a floor than a ceiling. Until an advisory resolves, federal civilian patching SLAs and procurement timelines for AI-aware vulnerability management remain undefined against a threat environment that has already materially shifted.

3 sources

Anthropic says Mythos can turn software patches into exploits in minutes - Axios
Exploit Evals - Anthropic Red Team
Measuring LLMs' impact on N-day exploits - Anthropic Red Team