Anthropic Red: Measuring LLMs' Impact on N-Day Exploits (Mythos Preview)
Source: Twitter/X (@logangraham, Anthropic) → red.anthropic.com/2026/n-days, 2026-06-08. Xiao, Abbott, Carlini, Cheng, Forsythe, Lucas, Nasr, Sakhuja. Raw: farmed
TL;DR
Anthropic's red team measured how much its upcoming Mythos Preview model accelerates N-day exploitation — attacking vulnerabilities that are already publicly disclosed and patched on some systems, where attackers race to hit machines still inside the "patch gap." Unlike zero-days (unknown to the vendor), N-days hand the attacker a roadmap: once a patch ships, "patch diffing" the pre- and post-patch binary reveals exactly what was fixed, and a working exploit becomes a matter of time. Historically that reverse-engineering was slow specialist work (WannaCry landed 59 days after the MS17-010 patch; the public Citrix Bleed exploit took ~2 weeks), which bought defenders time to roll out updates. The finding: Mythos Preview is meaningfully better at turning patches into N-day exploits — a few thousand dollars and a few hours of model time. Anthropic publishes this deliberately to give defenders lead time, with co-author @logangraham noting Mythos "will probably look trivial in a year" and promising more blue-team / defensive work.
Key points
- N-days may be the more dangerous class than zero-days precisely because the patch is a public blueprint; LLMs that compress patch-diffing collapse the defender's patch-gap window.
- Concrete cost numbers ("a couple thousand dollars and a few hours") make this a quantified capability measurement, not a vibes warning — the same evaluation-first posture as Anthropic's earlier cyber writeups.
- Disclosed pre-public-model-release. The point is to warn defenders before the capability is widely available, framing model-capability disclosure as a defensive act.
- Lands the same week as a flurry of Mythos public-release rumors on X (a "public version of Mythos tomorrow"; a "Claude 5 fable" variant with stricter cyber guardrails reportedly in 48h), so the cyber-capability framing is part of the launch narrative.
Relation to prior wiki state
- Extends the responsible-ai cyber-capability thread. Continues Anthropic's program of publishing offensive-capability measurements (zero-day work earlier) and pairs with the broader agentic-security backdrop: ChatGPT's Lockdown Mode (06-08, blocks exfiltration but not prompt injection) and OpenAI's agent-superapp tool access. Capability is being measured on the offensive side faster than defenses ship.
- Frontier-lab-as-its-own-red-team is now a recurring pattern (Anthropic Project Glasswing for AI-powered security appeared in the SemiAnalysis newsletter today). The defensive-tooling response is the thing to watch.
Gaps
The post reports a capability uplift but (in the surfaced excerpt) does not quantify the defensive counter — whether the same model accelerates patch deployment or vulnerability triage by a comparable factor. The "looks trivial in a year" framing is a projection, not a measurement.