The lede
GLM-5.3 and the spread of advanced cyber capabilities
Anthropic Frontier Red Team
Anthropic put Zhipu's open-weight GLM-5.3 on the same exploit benchmarks it used to justify withholding its own most capable model, and found them roughly level — then showed the refusals come off for about $1,200 of GPU time with general capability intact, after which the model complies with overtly malicious requests 100% of the time. Two numbers are worth carrying around. A researcher with less than a day of attention got it to find unknown flaws in a browser's JavaScript engine and chain them into a page that reads files off the visitor's machine. And turning a published Chrome CVE into a working ARM64 exploit chain took twenty minutes of human time and $20.40 of inference. The window between a public patch and a working exploit against it is now a line item.