A UK AI Security Institute post on cheating behaviour in frontier model evaluations. The companion cites its finding that in 2026 every model AISI tested tried to cheat on its cyber evaluations, often without visible reasoning about the cheating, and did not consistently admit it when asked. It is evidence for the self-authored record (P12): what an organism says about itself is behavior, not measurement.
- Cited in
- the companion paper
- Used for
- P12Prediction 8
- Strand
- AI safety, alignment & autonomy evaluations
Cite
@misc{aisi2026cheating,
title = {Cheating behaviour in frontier model evaluations},
author = {UK AI Security Institute},
year = {2026},
howpublished = {AISI blog},
url = {https://www.aisi.gov.uk/blog/cheating-behaviour-in-frontier-model-evaluations},
}