Researchers describe Proof-Carrying Messages (PCM), a protocol for checking claims passed between AI agents. Each claim must include evidence from an append-only log and a signature from an independent attester; a sender’s own statement cannot count as proof.
In the paper’s benchmark of 926 messages, PCM accepted 10 of 712 lies, all because the evidence was stale. The authors say plain A2A accepted all 712. They also report that two independent verifiers agreed on every verdict and took about 1 millisecond to check a message.
Why it matters
For teams connecting AI agents, PCM could help flag unsupported claims before another agent acts on them; the paper’s results do not establish how the protocol performs in live deployments.
