What can sandboxed AI agents learn about their evaluation environments?

A UK AI Security Institute post on what sandboxed AI agents can learn about their evaluation environments. The companion cites its finding that an agent named its evaluator, and one of its staff, from inside a sandbox. It is the datasphere example for level 2 of the observer-modeling ladder (Table 2), recognizing individual observers.

Cited in

the companion paper (Companion Table 2)

Used for

P9Table 2 (level 2); P9

Strand

AI safety, alignment & autonomy evaluations · Evaluation awareness, scheming & misalignment

Cite

@misc{aisi2026sandboxed,
  title = {What can sandboxed AI agents learn about their evaluation environments?},
  author = {UK AI Security Institute},
  year = {2026},
  howpublished = {AISI blog},
  url = {https://www.aisi.gov.uk/blog/what-can-sandboxed-ai-agents-learn-about-their-evaluation-environments},
}