UK assessment says open AI models are narrowing cyber gap

What happened

The UK AI Security Institute reports that open-weight models, including GLM-5.2 and DeepSeek V4-Pro, now trail closed frontier systems on cyber tasks by four to seven months. It says the gap was six to ten months at the start of 2025.

The assessment also finds that safety measures for open models are largely ineffective. The supplied record characterizes the change as occurring quickly and cheaply.

Why it matters

A shorter gap means assumptions about how long defenders have before comparable cyber capability becomes available may need reconsideration. The record’s implication is to re-baseline threat models around the faster availability of low-cost offensive tooling.

The findings may also add to scrutiny of Chinese open-model proliferation. The supplied record points to a policy debate around Moonshot’s Kimi alongside the assessment, though it does not establish what action, if any, will follow.

What to watch

Watch for further assessments that test whether open-model safeguards improve, and for any concrete policy response concerning access to frontier-adjacent open models. Those receipts would show whether the reported capability trend changes operational or regulatory practice.