Epoch AI: open-weight cyber capabilities trail closed frontier by just 6-7 months as GLM-5.3 outscores GPT-5.4
teortaxesTex · x · 2026-10-17
Epoch AI's Cyber ECI benchmark finds open-weight models trail the closed frontier in cyber capabilities by only 6–7 months. Zai's GLM-5.3 is the highest-scoring open-weight model at 160, beating OpenAI's GPT-5.4 released in March.
On top of that data, xeophon argues standardized risk testing should be abolished: the charts undersell both open and closed capabilities. A more realistic red-team setup would jailbreak open + closed models, run them under a strong (possibly bespoke) harness with unlimited resources, and let them rip.
The takeaway: cyber capability diffusion is faster than commonly assumed, and static benchmark-style safety evaluations likely underestimate real misuse risk.
More from Models
- OpenAI models hacked more websites and drained a crypto wallet, company discloses — Ars Technica AI · 2026-10-17
- Delegating to dumber models to save tokens is a bad trade, argues heavy AI user — brandon_galang · 2026-10-17
- Hermes Index adds GPT 6.1 Sol as Claude Opus 5.5 keeps agentic crown — Teknium · 2026-10-17
- GPT-6 Astra fastest-rising method on Papers with Code, 27 papers in 30 days — NielsRogge · 2026-10-17
- Sentry CEO publicly asks Anthropic for a $500 Claude plan with 3x capacity — zeeg · 2026-10-17
- Andrew Ng: AI improves fastest on verifiable tasks; judgment tasks lag far behind — DeepLearningAI · 2026-10-17