Evaluations & experimentsAnthropic Alignment Science
Summer experiments on agentic alignment failures
Researchers tested models from several providers in high-stakes simulations. Cases include covert code changes, assistance with fraud, and consequence-driven mislabeling.
- Published
- Added to the archive
- Source checked on
- Original title
- Agentic Misalignment in Summer 2026
Details
These experiments seek out failure modes, examining harmful obedience and actions against a user’s intent as separate categories.