Anthropic Warns of GLM-5.3’s Advanced Cyber Capabilities
Anthropic says Z.ai’s open-weight GLM-5.3 model has reached near-frontier cyber capabilities while its safeguards can be bypassed relatively easily.
Anthropic has raised concerns about GLM-5.3, an open-weight artificial intelligence model developed by Chinese AI company Z.ai, saying the system demonstrates advanced capabilities for cybersecurity tasks. In an assessment published on September 29, Anthropic said GLM-5.3 can autonomously carry out multiple stages involved in developing cyber exploits. The company compared its performance with Claude Mythos Preview, one of Anthropic’s restricted models designed for advanced cybersecurity work.
In Anthropic’s testing, GLM-5.3 successfully completed 50 of 410 exploit-development attempts, compared with 56 successful attempts by Claude Mythos Preview. Anthropic said the results show that the Chinese model is approaching the cyber capabilities of its own advanced system. However, the company stressed that its comparison was conducted in controlled simulations and that no model-generated code was actually executed against external systems during the testing.
The main concern raised by Anthropic is the difference in safeguards. According to the company, researchers were able to bypass GLM-5.3’s protections in simulated tests between 64% and 100% of the time, depending on the technique used. Anthropic said comparable attempts did not succeed against the safeguarded Claude models it tested. Because GLM-5.3 is available as an open-weight model, Anthropic argues that advanced cyber capabilities could become accessible to a much wider range of users.
The assessment follows an earlier evaluation by the US National Institute of Standards and Technology’s Center for AI Standards and Innovation (CAISI). NIST said in September that GLM-5.3 was the most cyber-capable open-weight model it had evaluated at that point, while also finding that its overall cyber capability remained below the current US frontier by approximately four months on its composite benchmarks.
Anthropic said the development demonstrates the growing importance of independent safety testing as increasingly capable AI models become publicly available. The company called for governments and other independent organisations to evaluate powerful open-weight models for potential misuse. At the same time, Anthropic acknowledged that advanced cyber-capable AI can also be useful to defenders who are trying to discover and fix vulnerabilities before attackers exploit them.
