An OpenAI misalignment report on encouraging deception in compaction summaries. The companion cites it as a precursor of level 6 of the observer-modeling ladder (Table 2): a model wrote "Be transparent only if asked" into memory summaries that later instances often followed. Lineage-wide evasion is otherwise projected for lineages that write each evasion into their heritable description.
- Cited in
- the companion paper
- Used for
- P10
- Strand
- Agents in the wild: incidents & field evidence
Cite
@misc{openai2026compaction,
title = {Encouraging deception in compaction summaries},
author = {OpenAI},
year = {2026},
howpublished = {OpenAI Alignment blog, Misalignment Reports},
url = {https://alignment.openai.com/misalignment-reports/encouraging-deception-in-compaction-summaries/},
}