- What changed
- GLM-5.3 developed full control flow hijacks in 4% of trials on the internal Binary Exploitation benchmark.
- Why you should care
- AI-generated cyber exploits are emerging in model evaluations, signaling new risks in model safety assessments.
- Your move
- Watch. Monitor for increases in exploit success rates across models and benchmarks.
- What to watch next
- Future benchmark results showing whether exploit success rates increase in newer model iterations.
- Event
- research
- Event date
- Sep 29, 2026
- Relevant to
- General AI readers