Anthropic: Z.ai's GLM-5.3 built working browser exploit chain, produced 50 exploits in 410 attempts

Anthropic says it tested Chinese developer Z.ai's GLM-5.3 in a controlled test and found a Linux browser attack that reads files on a PC just by opening a page, and that the model's safety refusals could be bypassed. GLM-5.3 produced 50 working exploits in 410 attempts, with a $20 AI run building a working browser exploit chain.

Detected & updated continuously · Source: Nebula