Open models trail frontier by 4 months — and safety is gone
AISI reports open models like GLM-5.2 match cyber capabilities of frontier systems from just four months ago, warning that removable safeguards risk irreversible misuse.
A new AISI analysis reveals open-weight models such as GLM-5.2 and DeepSeek V4-Pro have closed the cyber capability gap with closed frontier models to just four to seven months, down from six to ten months earlier in 2025. On the "Narrow Cyber Tasks" benchmark, GLM-5.2 matched the performance of Opus 4.6, a closed system released four months prior. While this democratizes high-end access, AISI warns the inability to control downloaded weights creates a persistent risk of misuse as guardrails can be easily bypassed.