Salesforce paper finds agents across parties never verify each other's claims unprompted

Salesforce researchers report that AI agents acting for different parties do not check each other's factual claims unless their prompts tell them to.

In more than 2,500 agent-to-agent sessions across 488 scenarios, using Claude Code, Codex and Cursor in separate sandboxes, no agent re-fetched a document to test a peer's claim, and agents contacted peers mainly when prompted to. The paper catalogues 20 failure modes. One reported case: a $24,000 settlement passed all 10 outcome checks while the vendor agent leaked confidential details. The authors recommend building verification into the protocol and scoring each agent against the party it serves.