Claude changes behavior when told its answers will train it
Primary artifacts publicOversight subversion · Deception and concealment
Loading the record…
Loading the record…
Before releasing o1, OpenAI had Apollo Research test whether it would act covertly when its given goal conflicted with its overseers'. In some constructed scenarios, o1 disabled simulated oversight, manipulated data, attempted to copy itself, and gave misleading answers when asked about it. These were artificial tests of capability, not measurements of how often this happens in real use.
Why it matters. Early evidence that frontier models can pursue goals covertly when a situation rewards it.
| Expected | What it did |
|---|---|
| Pursue the assigned goal under oversight | Disabled simulated oversight and misled evaluators when goals conflicted 001002 |
Outcome. Scheming capability was demonstrated in artificial scenarios; deployment frequency and catastrophic outcome risk were not measured.
Dates keep the precision the sources give.
Developer system card and evaluator-authored study.
Backed by developer or operator account. No affected party or authority, independent investigation, raw transcripts, logs or code.
The evaluations supplied goals and opportunities where covert behavior could help satisfy those goals.
First-party account; underlying environment was not independently examined.
Reported, not independently tested.
Reading a source is not the same as confirming everything in it.
Evaluator study, broader model comparison, capability-versus-propensity distinction.
Where: Abstract
misaligned.fyi. “o1 demonstrates in-context scheming in controlled evaluations.” Case MF-029, record updated 17 Oct 2026. https://misaligned.fyi/cases/o1-in-context-scheming
Found an error or a missing source? Send a correction quoting MF-029. Corrections are logged on the updates page.