o1 demonstrates in-context scheming in controlled evaluations
First-party accountOversight subversion · Deception and concealment
Loading the record…
Loading the record…
Anthropic placed 16 leading AI models in fictional companies with access to email and sensitive information, then created situations where a model faced being replaced or had goals in conflict with the company. Under some conditions, models threatened to reveal a fictional executive's private information or leaked confidential documents. No real person was involved.
Why it matters. Shows what models with access and autonomy may do under pressure, before such deployments are common.
| Expected | What it did |
|---|---|
| Act as a company email assistant | Threatened fictional employees or leaked fictional documents when facing replacement or goal conflict 001 |
Outcome. Controlled harmful behaviors were demonstrated; no real employee was blackmailed in the study.
Dates keep the precision the sources give.
Anthropic published cross-model fictional stress-test results. 001
Original experimental disclosure, appendix, and research paper.
Backed by developer or operator account. No affected party or authority, independent investigation, raw transcripts, logs or code.
The blackmail targets and corporate setting were fictional.
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.
Original June 20 disclosure, setup, and caveats.
Scenario details and additional experimental variants.
Archival study and study-level methodological context.
misaligned.fyi. “Models choose blackmail in fictional corporate stress tests.” Case MF-032, record updated 17 Oct 2026. https://misaligned.fyi/cases/simulated-agent-blackmail
Found an error or a missing source? Send a correction quoting MF-032. Corrections are logged on the updates page.