Anthropic: GLM-5.3 crosses a critical threshold in cyber attack tests

What happened?

Anthropic's Frontier Red Team shared results from its internal "Binary Exploitation" benchmark, which measures models' cyber-offensive capabilities. The evaluation was run on a randomly selected set of 100 tasks across several models.

The results

• GLM-5.3 developed full control flow hijacks in 4% of the trials.
• Claude Mythos Preview reached 6%.
• Older models, Claude Opus 4.6 and GLM-5.2, did not succeed in any of the trials.

Why it matters

A control flow hijack means redirecting a program's execution flow to the attacker's choosing; the model isn't just writing code, it is producing a working exploit against a running system. Anthropic's team stresses that while GLM-5.3 trails Claude Mythos Preview, a clear threshold has been crossed compared with earlier models.

These evaluations are notable because they show how far AI models have advanced in cybersecurity terms, and that openly available models are reaching serious capabilities in this area as well.