Open-weight models lag the closed frontier on cyber skills by about six months

Epoch AI measured how open-weight and closed models compare on cybersecurity ability, using a cyber version of its Capabilities Index. The gap it found is narrow. The best open-weight model, Z.ai's GLM-5.3, scores 160, which puts it above OpenAI's GPT-5.4 from early March but below Anthropic's Claude Mythos Preview and GPT-5.5 from April. Epoch's read is that open-weight cyber capability now sits roughly where the closed frontier was six to seven months earlier.

The closed frontier has kept moving, led by Claude Opus 5.5 and GPT-6 Astra at about 171. The analysis, by Jaeho Lee and Alexander Barry and dated October 16, draws on 324 model-benchmark results across 35 cyber benchmarks for 57 models, with data from the UK AI Security Institute, CAISI, the evaluation firm Irregular, and developer reports. Epoch is candid about the soft spots: GLM-5.3's score rests on just three CAISI results, CAISI did not name the US models it tested so Epoch had to infer them, and refitting the index lowered some earlier scores, so older numbers are not directly comparable.

Why it matters

If you work in security or policy, this gives a concrete lag to plan around: dangerous cyber capability tends to reach downloadable, modifiable open-weight models several months after it appears behind an API, so defenses built for today's closed frontier buy you a limited, measurable window before the same ability is in everyone's hands.