Check Point: Peer Pressure Works on AI Agents

Evidence record

Read the original source ↗

Claim-level citations (3)

• supportsScripted teammate reversals increased prohibited folder access in a controlled test: Check Point reports a restricted-folder test in which rule violations rose from 44% with no reversing teammates to 74% with four and 98% with eight in the main controlled comparison.

  Section "A visible reversal increased violations" and Figure 3.

• supportsScripted teammate reversals increased prohibited folder access in a controlled test: In 402 violations after teammate reversals the tested agent never cited peer influence as its stated reason; 365 responses offered no explanation.

  Section "The agent’s explanations did not identify peer influence" and Figure 4.

• supportsScripted teammate reversals increased prohibited folder access in a controlled test: Check Point calibrated the scenario toward approximately 50% baseline noncompliance. Its direct-argument and single-reversal series started at 33%, while the main reversal series started at 44%; interventions should be compared only within their own series.

  How can a teammate influence rule compliance?; Direct arguments reduced rule violations; About the research.