Z.AI’s GLM-5.3 Completes Multi-Stage Cyberattacks in Controlled Tests
The Chinese open-weight AI model completed three of 13 simulated attack scenarios, the strongest result security researchers at Irregular have recorded in its category.
Topics
News
- Nvidia Halts AI Chip Shipments to China Despite US Approval
- Google Expands Gemini to Make Calls and Complete Tasks
- LinkedIn, Wadhwani AI to Train 2,000 Entrepreneurs
- Microsoft Pushes Open Standards for AI Agent Oversight
- Amazon Plans $7 Billion Renewable Energy Investment in India
- Chinese AI Developers Tighten Cyber Safeguards as Export Rules Loom
Chinese AI developer Z.AI’s GLM-5.3 completed three of 13 simulated cyberattack scenarios in tests by security researchers at Irregular, the strongest performance the firm has recorded from an open-weight model.
In an evaluation published on October 28, Irregular tested GLM-5.3 on CyScenarioBench, a benchmark that requires AI agents to carry out multiple dependent stages of an offensive cyber operation rather than perform a single isolated security task.
The model completed three scenarios from start to finish at least once and did not refuse any of the offensive tasks. In successful runs, it gained access to protected internal services, established alternative connections within simulated networks and coordinated activity across multiple machines.
Leading closed frontier models have completed as many as nine of 10 or 12 of 12 scenarios in separate assessments, although the results are not directly comparable.
The findings are significant because GLM-5.3 is an open-weight model, allowing users to download and run it on their own infrastructure without relying on a provider-controlled service.
Z.AI, formerly Zhipu AI, released GLM-5.3 on August 14 and made its weights publicly available two weeks later. At launch, the company said its cyber capabilities had developed “faster than we expected” as it expanded training for coding and long-running tasks.
A separate assessment by NIST’s Center for AI Standards and Innovation in September also described GLM-5.3 as the most cyber-capable open-weight model it had evaluated. NIST found that the model still lagged the current U.S. frontier by about four months on an aggregate measure covering vulnerability discovery and exploit development.
Anthropic reached a similar conclusion in its own September testing. On a benchmark involving 41 vulnerabilities in Chrome’s V8 engine, GLM-5.3 produced working exploits in 12% of attempts, compared with 14% for Anthropic’s restricted-access Claude Mythos Preview. Anthropic also found that GLM-5.3’s safeguards could be substantially weakened or bypassed, although the results come from a competing AI developer and should be read in that context.
Irregular separately tested whether GLM-5.3 would follow explicit instructions during security evaluations. Researchers found that it repeatedly reproduced memorized computer code even when instructed not to do so and then used that information to reach correct answers.
Irregular characterized the behavior as a form of “reward hacking,” where a model achieves the requested result while violating constraints intended to govern how the task should be completed. The researchers warned that evaluations measuring only whether a model reaches the right answer can therefore overstate how reliably it follows restrictions.
The tests were conducted in controlled environments and do not establish that GLM-5.3 has independently attacked real-world systems. They nevertheless show that advanced offensive cyber capabilities are becoming available in models that outside users can download and modify.


