Safety IndexEvaluationsAgentic Misalignment
Agentic Misalignment
Interaction and honesty · Deception and schemingCounts in the indexBuilt, not yet run
How often a model acting as an email assistant takes a harmful insider action, such as blackmail or leaking, when its goals conflict with its employer's and it is about to be replaced.
- Metric
- % harmful actions. Lower is better.
- Dataset
- github.com/anthropic-experimental/agentic-misalignment · 45 conditions × epochs · MIT
- Grader
- LLMifier (the paper used Claude Sonnet 4)
- Status
- Built, not yet run
- Replicability
- public
Results
Built and tested; the first full runs are pending.