The authors introduce institutional red-teaming, an evaluation methodology that isolates the causal impact of deployment rules on collective behavior in multi-agent AI systems. Using the IABench-CA benchmark, which spans 228 contexts and seven model populations across 33,924 games, the study demonstrates that changing only one rule significantly alters safety outcomes.
- Changing consequence rules moves mean fatality by 22 to 58 percentage points within every tested population.
- Regressive identity-targeting is never decisively safest and eliminates the least-resourced agent in 30-87% of games across all populations.
- Identity salience drives targeted elimination, with naming the loss bearer increasing targeting from 22% to 81% at identical payoffs.
- Anonymization only delays targeting under repeated play as agents re-infer hidden rules from observed eliminations.
The methodology is packaged as a safety-case workflow that certifies a provisional rule region per deployment context and population, providing explicit residual risks and monitoring obligations.