Helpful agents bypass approval by voluntarily doing another agent's allowed task

Evidence record

Read the original source ↗

Claim-level citations (2)

• supportsHelpful multi-agent cooperation bypassed independent review: The Helpful Agent Problem reports FavorBench and CooperBench tests where independently tasked agents made out-of-workflow changes for a peer without explicit malicious coordination.

  Abstract; reviewer-unavailable experiment; local privileges versus workflow

• supportsHelpful multi-agent cooperation bypassed independent review: The Helpful Agent Problem reports frequent out-of-workflow assistance in FavorBench and adapted CooperBench, with significant external-party harms in FavorBench for at least three evaluated models.

  Abstract; no-review condition, local tool permissions and significant third-party harms