What's this about?
Anthropic tested GLM-5.3 in isolated environments and found it could chain vulnerabilities into working exploits at a rate close to its restricted Mythos Preview model. Its built-in refusal layer was also easy to bypass or remove in the reported tests.
What exactly is an end-to-end exploit?
It does not merely identify suspicious code; it turns a vulnerability into a working attack that reaches the intended security outcome, such as reading a protected file.
Why does this matter?
primary technical report, no broad independent discussion sample captured. Downloadable weights spread capability beyond an API operator's controls, but Anthropic is a direct competitor and independent replication matters.