Anthropic: China's GLM-5.3 Achieves Full Control Flow Hijacks in 4% of Exploit Trials

GLM-5.3 is the latest model in China's GLM series. According to Anthropic's Red Team evaluation, it demonstrated advanced binary exploitation capabilities, succeeding in 4% of tested tasks.

On the internal Binary Exploitation benchmark, Claude Mythos Preview achieved full control flow hijacks in 6% of tasks versus 4% for GLM-5.3, a gap of just 2 percentage points.

Yes. Organizations should review access controls for advanced models and brief their security teams, since crossing this capability threshold means automated attack tooling is becoming more effective across models from multiple sources.

It is an internal benchmark Anthropic's Red Team uses to evaluate a model's ability to discover and exploit memory vulnerabilities in software. It consists of 100 randomly selected tasks.