Anthropic: GLM-5.3 exploits near Claude Mythos Preview

Anthropic's report lands in a debate that was already running. The UK's AI Security Institute recently found that open models had narrowed their cyber-capability lag from six to ten months down to four to seven months, and warned that their safeguards are largely ineffective. What was unclear then was whether open models would also catch up to the leap that Claude Mythos Preview represented. The ExploitBench and internal binary exploitation measurements from Anthropic, plus CAISI's independent assessment, suggest GLM-5.3 is a first answer to that question.

Anthropic frames its own release choices as the counterfactual. It deliberately held Mythos Preview back, giving access only to select defenders through Project Glasswing, who the company says have since found more than 10,000 vulnerabilities in critical software; OpenAI is taking a similar approach with Daybreak. GLM-5.3, by contrast, is available for anyone to download, and Anthropic says several developers released unlocked versions within days of its launch. Its abliteration experiment showed how quickly refusals can be stripped, though the simulation does not execute code and so cannot show whether an attack would actually have succeeded.

The report is not a neutral document. Anthropic does not release its model weights and presents that as a security advantage, and a cheap Chinese open-weight model close to the frontier is a direct competitor; its call for government testing of GLM-5.3's successors also invites suspicion of regulatory capture. Still, the capability numbers are backed by CAISI, and unlocked versions are already out there, so the practical question for defenders and governments is likely to be less whether the gap has narrowed than how fast they can prepare while it still exists.