GLM-5.2 matches closed models on cyber tasks
The UK AI Security Institute (AISI) has released evaluation results from testing leading open-weight AI models against closed frontier systems on practical cyber work, such as vulnerability research, reverse engineering, exploitation, and multi-step network attacks. The benchmark results indicate that the performance gap between open-weight and closed-weight models is shrinking rapidly, with Z.ai's open-weight GLM-5.2 matching the cyber capabilities of closed frontier models released just four to seven months prior.
Open-weight models are no longer second-tier options for advanced cognitive tasks; their rapid parity with proprietary models will permanently disrupt the economics of the frontier AI market and complicate global regulatory efforts.
- –The capability gap between leading open-weight models and closed-source giants has shrunk from 6-10 months in 2025 to a mere 4-7 months in mid-2026.
- –GLM-5.2's performance on multi-step network attacks and vulnerability research indicates that highly capable, dual-use offensive cyber tools are now freely downloadable under permissive licenses.
- –Because open-weight model safety guardrails can be bypassed through fine-tuning, the proliferation of frontier-class open weights presents a persistent security challenge that post-release controls cannot mitigate.
DISCOVERED
2h ago
2026-07-18
PUBLISHED
2h ago
2026-07-18
RELEVANCE
AUTHOR
aispace_bot