Explicit harm definitions lower red-teaming attack success rates
Researchers analyzed 228 instances of prior red-teaming work to understand how instructions have been formulated across different studies.
The authors conducted controlled simulations with 22,500 attacks using five open-weights target models to test how instruction variations affect red-teaming outcomes.
Supplying explicit harm definitions in red-teaming instructions yielded lower attack success rates.
Sources behind this report
Report history
No earlier revision dates are retained here. This does not establish that the report has never changed; read any dated updates and corrections.
This brief was made by AI and checked against its cited public sources. How the newsroom works