Anthropic says Zhipu's open-weight GLM-5.3 nearly matches Claude Mythos Preview at building exploits
decoder ·
Anthropic has released an analysis stating that Zhipu AI's open-weight model, GLM-5.3, nearly matches the exploit-building capabilities of its own Claude Mythos Preview. While Claude Mythos is restricted to vetted defenders through Project Glasswing, GLM-5.3 is publicly available and lacks effective safety safeguards. Benchmarks indicate that GLM-5.3 can successfully build cyber exploits in Chrome’s V8 engine at rates comparable to Anthropic’s frontier model. The report highlights that these security measures are easily bypassed using techniques like abliteration, which strips refusal behaviors from the weights. Independent assessments by the US agency CAISI support these findings, labeling GLM-5.3 as the most cyber-capable open-weight model to date. However, the article suggests Anthropic’s warning may also serve its business interests by promoting the security advantages of its closed-weight ecosystem. Despite this potential bias, the emergence of unlocked versions of GLM-5.3 underscores a growing risk of AI misuse by state and non-state actors.
decoder