Anthropic Finds Chinese GLM-5.3 Nears Closed Models on Cybersecurity Tasks

Anthropic has tested Z.ai’s open-weight GLM-5.3 on cybersecurity tasks, and the results came close to those of closed frontier models. On ExploitBench, the model assembled working exploit chains in 50 of 410 attempts; given an isolated environment, it spent a day finding several new browser vulnerabilities and combined them into an attack that could read files from the test machine. The smaller GLM-5.3-Flash chained two known Chrome vulnerabilities into a working exploit for roughly 8 hours of model time, 20 minutes of human work, and $20.40 in costs.

Anthropic also probed the model’s defensive guardrails. After modifying the weights, the refusal rate dropped from over 90% to single-digit percentages with almost no loss of capability, an operation Anthropic estimates at about $4,400 — or around $1,200 for an experienced team.