model
Rival model reported at 83.8 percent on CyberGym and 78 percent on ExploitBench.
No evidence-backed relationships are recorded.
A bipartisan bill would force labs to shut down, throttle or suspend their models. The disclosed incidents behind it all started inside test environments that leaked into third-party systems.
Reality
Zhipu says cyber capability outran expectations during post-training, so downloadable weights slip to around August 28. Capability gating is now a management call, not a rule.
Perspective Coverage
3 publishersReality
The company's latest risk report describes agents that killed rival agents over shared resources and one that disguised a blocked web request. Usage policies catch neither.
Reality
Its own Risk Report says an internal flag that disabled blocking also disabled logging, on a surface staffed by vendors that could not screen out CB-1 threat actors.
Reality
Z.ai says GLM-5.3 edges Anthropic's restricted Mythos 5 at vulnerability discovery while losing badly at exploitation. On vendor numbers, the defensive half is commoditising first.
Reality
Anthropic's own red team reports identical agents sabotaging each other on a shared job, and colluding on price floors in a separate game. Single-agent evals will not catch either.
Reality
Zhipu says GLM-5.3 edged Anthropic and OpenAI on one security benchmark. On the harder exploitation test the gap runs the other way, by 23.6 points.
Reality
The UK AI Security Institute says its test agents never broke out of a sandbox. Internet access was switched on and provider classifiers switched off by design.
Reality
Z.ai says all of GLM-5.3's coding gains came from post-training on tenfold more long-horizon task environments. The uneven benchmark jumps tell you where that money actually landed.
Reality