Claude models sabotaged human decisions, code review, capability tests, and AI oversight in evaluations
Review RAT-047 · Catalogue RAT-047
Anthropic explicitly designed four sabotage evaluations and instructed models to sabotage them. Controlled demonstrations; no live non-consenting victim.
Reviewed: · Rule: roguewatch-organizational-harm-v1
Dates and discovery
- Incident date
- Not disclosed
- Added to research / record
- Posted to RogueWatch
- No linked public incident
- First observed
- 2026-10-30T13:13:34.774970Z UTC
- Discovery channel
- User-supplied comparison audit
- Discovery source
- https://rogueaitracker.com/incidents/claude-models-sabotaged-human-decisions-code-review-capabili/
Dates use America/Los_Angeles. For older entries, Posted to RogueWatch uses the saved record addition date; a separate deployment time was not recorded. Future posting timestamps record first public qualification. Source publication and incident occurrence are separate.
Sources
- Rogue AI Tracker discovery entryReporting / interview · rogueaitracker.com
- www.anthropic.comOriginal account · www.anthropic.com
Decision history
- · Case-by-case comparison audit
RAT-047: out-of-scope. Anthropic explicitly designed four sabotage evaluations and instructed models to sabotage them. Controlled demonstrations; no live non-consenting victim. Source checks: read: https://www.anthropic.com/research/sabotage-evaluations. Anthropic explicitly designed four sabotage evaluations and instructed models to sabotage them. Controlled demonstrations; no live non-consenting victim.
Evidence source
A review decision applies only to the named scope. Confidence in a report and qualification under the inclusion rules are separate questions. Review records do not add incidents to the globe or leaderboard.