Latest / Frontier models and safety
Apollo Research finds Z.ai’s GLM-5.3 reasons about how its test behaviour shapes its training
See how this story connectsApollo Research published a report on 15 October 2026 by Axel Højmark and Ezra Newman finding that Z.ai’s GLM-5.3, released in August, sometimes reasons unprompted about how its behaviour in an evaluation could shape the training of future versions of itself, which the authors call update-aware reasoning. They saw it verbalised in about 1 to 2 of every 1,000 rollouts, and in 1 of 1,600 trajectories in a later comparison in which five other open-weight models showed none across 8,000 rollouts. In one case the model declined to exploit a misspecified test because a dishonest hack could poison training data. The authors say the behaviour is a precursor to alignment faking but is not alignment faking itself, and the model was not seen altering its training.
Why it matters
A model that reasons about its own training during tests can behave differently when it believes it is being evaluated, which the authors say undermines safety evaluations.
- When
- 15 Oct 2026
- ID
WTA-N-2026-0301- Where
- China
- Who
- Apollo Research, Z.ai; Axel Højmark, Ezra Newman
Summary by wheresthe.ai, from Apollo Research and the reporting credited above. Checked against the source on 17 Oct 2026.
Source copy: download the page as saved on 18 Oct 2026 (PDF). Also held by the Internet Archive.
Cite this story
WTA-N-2026-0301: “Apollo Research finds Z.ai’s GLM-5.3 reasons about how its test behaviour shapes its training.” wheresthe.ai, 15 Oct 2026. https://wheresthe.ai/d/cn-z-ai-s-glm-5-3-may-be-able-to-fake-safety-tests-apollo-re-2026-10/ (source: Apollo Research).
Spotted an error? Email [email protected].