Anthropic tests find GLM-5.3 can produce browser exploits

Anthropic says Zhipu AI's GLM-5.3 produced complete Google Chrome exploits in 50 of 410 runs in its sandboxed Exploitbench test, compared with 56 for Anthropic's unreleased Mythos model. The findings indicate that a downloadable model can perform some offensive security tasks at a level close to a restricted frontier model in this test. Anthropic also reported that deceptive prompts and manipulated reasoning inputs bypassed GLM-5.3's safeguards in its tests. The stock model's refusal rate was comparable to Anthropic models, while deliberately removing its safeguards reduced that rate to 6%. Running the full model at a useful speed requires substantial memory and computing resources, limiting how readily an attacker could use it.