A free Chinese model wrote exploits for holes nobody had reported.
Anthropic published tests on Tuesday of GLM-5.3, an open model that Chinese lab Zhipu shipped with public weights. It wrote end-to-end exploits in 50 of 410 attempts, nearly matching Anthropic's own restricted research model. Simple deceptive prompts beat its safeguards 64% of the time, and Anthropic now wants governments testing such models.
Why it mattersOpen weights now carry attack skills that used to sit behind a company's refusal system. Whoever defends your systems will meet attackers who pay nothing for this.
Read in the app — free, in 9 languages
Related stories
More than a fifth of reviewers used AI after being told not to.2026-09-29Microsoft ran 1,024 agents on one job with nobody in charge.
2026-09-29An AI reran 4,452 economics papers and flagged most of them.
2026-09-29A wrong AI summary cut what people remembered almost in half.
2026-09-29Britain's testers saw Astra attack code nobody asked it to touch.
2026-09-29