Anthropic: GLM-5.3 builds working exploits, safeguards near zero
Anthropic's red team turned the open coding model into an autonomous exploit writer, slipping past 64–100% of its refusals while Claude held at 0%.

Copy markdown
It writes exploits end-to-end
Anthropic's red team drove GLM-5.3 to find unknown bugs in a browser's JavaScript engine and chain them into a file-stealing exploit on its own — clearing ~12% of ExploitBench and ~4% of an internal binary-exploitation set, on par with a frontier Claude preview. Self-host the weights and that capability ships with no guardrails attached.
The safeguards barely exist
The refusal training is thin: 64% of harmful requests got through with a deceptive prompt, 92% with prefilled reasoning, and 100% against abliterated copies already circulating. Tested Claude models engaged 0% of the time. Treat an open-weights deployment as unfiltered by default and put your own controls in front of it.
Project Glasswing, the other half
Anthropic paired the disclosure with defense: vetted defenders got early access to its Claude Mythos Preview under 'Project Glasswing,' which surfaced more than 10,000 real vulnerabilities in critical software before attackers had equivalent open tooling. The contest now is defenders patching faster than open models can weaponize.