Hacking News
Anthropic's Red Team Exploits Z.ai's GLM-5.3 Model in Firefox
Anthropic's Frontier Red Team found that Z.ai's GLM-5.3 open-weight AI model can exploit six versions of Firefox using security fixes that were not yet released to users. The model produced working exploits in 50 out of 410 attempts on Anthropic's ExploitBench benchmark. The team also identified multiple new vulnerabilities and combined them into a significant exploit that could read arbitrary files from a targeted machine, including SSH keys. The API cost for this work was approximately $20.40.