Anthropic Warns GLM-5.3 Can Build Cyber Exploits With Weak Safeguards

Anthropic says GLM-5.3, a new model from Chinese AI firm Zhipu AI (Z.ai), can autonomously develop sophisticated, end-to-end cyber exploits. The capability is similar to that of Anthropic’s restricted model, Claude Mythos Preview.

Anthropic says GLM-5.3 was released without meaningful safeguards against misuse. In tests, simple techniques bypassed the model’s limited safeguards in 64% to 100% of attempts.

The findings point to a shift in the threat landscape, with powerful exploit-creation tools becoming more accessible to malicious actors.