safety security
Remove the Few Messages That Enable the Violation
A necessity-and-sufficiency method cut controlled multi-agent violations from 93 percent to 6 percent with fewer deletions.
Summary
A necessity-and-sufficiency method cut controlled multi-agent violations from 93 percent to 6 percent with fewer deletions.
Collective Dependence Analysis tests which individually harmless messages were necessary and sufficient for a group of language-model agents to produce a policy violation. Across constructed cases, unscripted conversations and recorded interactions, selected removals reduced violations from 93 percent to 6 percent in controlled settings. The method used a linear number of model evaluations and matched that reduction with half as many deletions as a frontier-model transcript judge in eligible cases. Unfamiliar tasks still produce incomplete explanations, and removing a message can disrupt correct behavior.
Why it matters
A necessity-and-sufficiency method cut controlled multi-agent violations from 93 percent to 6 percent with fewer deletions.
Limits and context
No additional limitation was separately recorded.
Key claims
A necessity-and-sufficiency method cut controlled multi-agent violations from 93 percent to 6 percent with fewer deletions.
Evidence: source-2026-11-10-004
Sources
- arXiv preprint 2611.11724, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.