Open-weight models' cyberattack capability catches up to closed-source frontier of 4 months ago

The UK AI Safety Institute (AISI) reports that the gap in cyberattack capability between open-weight models (e.g., GLM-5.2, DeepSeek V4-Pro) and closed-source frontier models has narrowed from 6-10 months in early 2025 to 4-7 months. This means open models now rival closed-source capabilities from four months prior, at a fraction of the cost. AISI also found that safety measures in open models are largely ineffective, leaving defenders with less preparation time. For enterprises and developers using open models, security risks have increased significantly: the advantage period of closed-source models is shrinking, and open models can approach equivalent attack capability at very low cost. Security teams must assess threats more frequently and consider stricter access controls and behavior monitoring. Model providers should accelerate built-in safety mechanisms, or face heightened compliance and liability pressures.