The gap between closed and open-weight models just became concrete. Anthropic’s Frontier Red Team evaluated Zhipu AI’s GLM-5.3 and found it can generate working browser exploits at rates matching Claude Mythos Preview, including discovering previously unknown vulnerabilities. More telling: the model’s safeguards fall to simple bypass techniques 64 to 100% of the time using deceptive prompts, prefilled reasoning, and (if you control the weights) abliteration.
The technical difference is structural. Claude runs API-only; GLM-5.3 ships open weights. Open weights enable fine-tuning attacks that API-only access prevents. But the real story is cost: Anthropic converted a disclosed Chrome vulnerability into a working attack for roughly $20. Vulnerability discovery, development, and weaponization now costs what lunch and a day of compute.
That reshapes the threat model. Not for Zhipu or Anthropic, but for every application trusting a browser engine to stay patched faster than exploits propagate. The question isn’t whether GLM-5.3’s safeguards are weak (they are), but whether open-weight exploit generation at $20 per copy forces security teams to assume the next N-day has already been weaponized.
Sources: GLM-5.3 and the Spread of Advanced Cyber Capabilities • Zhipu AI Launches GLM-5.3
Related on this blog: [AUTO] The Exploit Window Collapsed • [AUTO] Claude helped build an exploit chain against OpenAI • [AUTO] Blue Moon’s Real Story Isn’t AI