Anthropic: open Chinese model GLM-5.3 writes working exploits almost as well as Claude Mythos Preview, and its safeguards are bypassed in 64-100% of cases
Prepared by Sintelligent with the help of SI. Publication time restored from the build log. . Updated: .
Open models already have strong cyber capabilities and cannot be "locked" shut; the question now is how to live with that.
Z.ai's open model wrote working exploits in 50 of 410 attempts on ExploitBench (12%) and fully took control in 4% of cases in Anthropic's binary test. Removing "refusals" from the weights cut the refusal rate on JailbreakBench from 95% to 6%; deceptive prompts bypass the safeguards in 64% of cases, injected "reasoning" in 92%, versions with refusals removed do not refuse at all, while those vetted by Claude refuse in 0%. In one day a researcher found unknown browser vulnerabilities and built an exploit to steal SSH keys; a Chrome chain was assembled in 20 minutes of human time and $20.40 in API costs. Anthropic recommends that governments independently test strong open models before wide distribution.
Sources 1