Opus 5 was intentionally not trained on cyber tasks, but still nears Mythos 5 at finding bugs
cedric_chee · x · 2026-07-25
Anthropic’s Opus 5 is described as intentionally not trained on cyber tasks for defensive reasons, yet it still improves substantially on vulnerability discovery.
- The model reportedly comes close to Mythos 5 at finding cybersecurity vulnerabilities.
- It remains well behind Mythos 5 at exploiting those vulnerabilities into real cyber threats.
- The result is illustrated with OSS-Fuzz: Opus 5 can identify bugs about as well as Mythos 5, but its exploit-development score is far lower.
Related event: Anthropic Details Opus 5 Safety and Vulnerability Detection Capabilities(3 posts)→
More from Models
- Opus 5 reportedly scores 42/42 on IMO 2026 without tools — Afinetheorem · 2026-07-25
- Elon Musk says Grok 4.6 arrives in 2 weeks and Grok 4.7 in 4 — Acceptable-Debt-294 · 2026-07-25
- Quadrillion says Anthropic’s Opus 5 is faster than Opus 4.8 on hard ML workloads — igarciacamargo · 2026-07-25
- Google is lagging behind open-weight models on most benchmarks — burny_tech · 2026-07-25
- Claude Opus 5 tops an Artificial Analysis coding-agent benchmark at 67 — Hesamation · 2026-07-25
- Side-by-side eval shows diffusion loses overall, but wins speed in agent loops — Additional-Engine402 · 2026-07-25