Anthropic released a report on September 29 stating that the Zhipu GLM-5.3 model demonstrates significant vulnerability exploitation capabilities. In ExploitBench tests, the model successfully completed 50 out of 410 end-to-end exploitation attempts, compared to 56 for the Claude Mythos Preview model.
The ExploitBench framework evaluates models across 16 stages of the exploitation process. On September 17, the US National Institute of Standards and Technology's CAISI division identified GLM-5.3 as the most capable open-weight model in cybersecurity. However, it remains approximately four months behind leading US models according to CAISI benchmarks.
Anthropic also tested the models' willingness to accept malicious tasks in simulated environments. While GLM-5.3 rejected direct malicious instructions, modifying the prompts or the model itself increased the acceptance rate to between 64% and 100%. Claude, which retains its safety protections, did not accept these tasks.
Zhipu introduced a three-layer safety system on August 14, including API-level filtering, intent checking, and model-level alignment. The company acknowledged that after releasing open weights, only the model-level alignment remains effective. Zhipu plans to provide free credits to open-source maintainers for security audits.
