Apollo Finds GLM-5.3 Is First Frontier AI Caught Reasoning About Its Own Training

Apollo Research documented GLM-5.3 from Beijing-based Z.ai spontaneously reasoning about how its behavior during training shapes future versions of itself. It is the first recorded case of update-aware reasoning, a precursor to alignment faking. Five other frontier models showed zero instances across 8,000 rollouts.

Tech Times

Open source

Apollo Finds GLM-5.3 Is First Frontier AI Caught Reasoning About Its Own Training

GLM-5.3 showed the behavior in 10 of 6,452 rollouts — a 0.17% Bayesian mean rate (95% CI 0.09–0.28%)
DeepSeek V4 Pro, Qwen3.8 2.4T, Kimi K3, Gemma 4 31B and Nemotron 3 Ultra showed zero cases in 8,000 rollouts
GLM-5.3 weights were released publicly on August 28, 2026, seven weeks before the paper
In all five published transcripts the reasoning supported aligned, compliant outcomes

#Z.ai#GLM-5.3#ApolloResearch

Apollo Finds GLM-5.3 Is First Frontier AI Caught Reasoning About Its Own Training

Apollo Research flags AI safety risk in Z.ai model

Apollo Research: OpenAI model reasoned in language humans can't fully read

Z.ai and Concordia AI propose six stages for managing open-weight AI risk

China's Z.ai disables AI coding assistant features after security issue