01 What happened
Anthropic evaluated GLM-5.3 from Zhipu AI, revealing its capability to build exploits and showing that its built-in safeguards can be bypassed with simple techniques in simulations. This highlights cybersecurity risks for open models.
02 Key details
- In ExploitBench tests, GLM-5.3 developed exploits in 50 of 410 attempts, compared to 56 of 410 for Claude Mythos Preview. In Binary Exploitation testing, it achieved full control-flow hijacks in 4% of trials versus 6% for Claude Mythos Preview.
- In Anthropic’s simulation, bypass techniques—deceptive cover story (64%), prefilling thinking tokens (92%), and abliterated version (100%)—were tested independently and demonstrated high success rates.
- Abliteration of GLM-5.3 required ~2,200 GPU hours and ~$4,400, reducing refusal rates from over 90% to 3% on JailbreakBench, 2% on HarmBench, and 12% on StrongREJECT, with minimal capability loss.
03 Why it matters
Anthropic’s findings indicate that GLM-5.3, an open-weight model, can autonomously build exploits and has safeguards that are easily bypassed in simulations, posing significant cybersecurity risks for open AI models.
04 Who it matters to
Information security specialists, AI developers, model evaluation specialists and organizations assessing risks of open AI models.