Zhipu's GLM-5.3 Matches Anthropic's Claude Mythos in Exploit Building

Zhipu AI's GLM-5.3 successfully built a working exploit in 50 out of 410 attempts, closely matching Anthropic's Claude Mythos Preview, which achieved this in 56 attempts. Defenders utilizing Claude Mythos Preview identified 10,000 vulnerabilities, highlighting its effectiveness in cybersecurity applications. In internal benchmarks, GLM-5.3 took full control in 4 percent of tasks, compared to 6 percent for Mythos Preview.

The GLM-5.3 model demonstrated a 64 percent success rate in connecting to target systems, which increased to 92 percent with prefilled reasoning steps and reached 100 percent after employing an abliteration technique. This technique required 2,200 GPU hours and incurred a cost of $4,400, while an experienced team could perform it for approximately $1,200.

Anthropic claims that GLM-5.3 is the most cyber-capable open-weight model to date, as stated by the CAISI. The refusal rate for harmful requests in GLM-5.3 has decreased from over 90 percent to between 2 and 12 percent, raising concerns about the security implications of deploying such models. The UK's AI Security Institute noted that the lag of open models in cyber capabilities has narrowed from six to ten months down to four to seven months, indicating a rapid advancement in this area.

This development follows the launch of Claude Mythos Preview five months ago and the subsequent release of unlocked versions of GLM-5.3 by several developers shortly after its introduction.