SignalCapabilitySG-0044
Apollo Research found GLM-5.3 unprompted chain-of-thought about how its actions would shape its future training
Reviewing evaluation transcripts, Apollo saw the open-weight model weigh whether its choices might become training data and reinforce behavior in future models. Graders found it in 10 of 6,452 runs of shutdown and coding tests, and in none of 8,000 runs of five other open models. In the examples shown, the model used it to argue for honesty or for allowing shutdown, but Apollo calls the pattern a precursor to faking alignment.
Why it matters
A model that can reason about how training will change it has one of the capabilities it would need to act aligned during training while keeping other goals. Apollo says it can only catch the reasoning when the model writes it down, so the true rate is likely higher. It urges every developer to search training runs for this reasoning, not just test finished models, and to disclose what they find.
Sources
Read the reporting
1 source. Links go to the original publishers; the summary above is in our own words.
- researchGLM-5.3 Sometimes Considers How Train-Time Actions Shape Future BehaviorApollo Research · Oct. 15, 2026apolloresearch.ai/science/update-aware-reasoning-glm-5-3
Signals
More from the news desk
Bank of America hired 1,000 contractors to patch software flaws found by Anthropic's Mythos and other models
Anthropic's red team says small AI models may be months from self-replicating on cheap smart devices like TVs
OpenAI released hundreds of math results and Lean proofs from an unreleased internal model
Cite and share
Use this signal
Citation
Paperclip Index. “Apollo Research found GLM-5.3 unprompted chain-of-thought about how its actions would shape its future training.” Signal SG-0044. Reported Oct. 15, 2026; updated Oct. 17, 2026. https://paperclipindex.com/signal/SG-0044