New PCM protocol lets agents accept only 10 of 712 incorrect claims

Researchers published PCM, a protocol in which each claim between AI agents carries proof. The agent that gets a message has only the report of the sender. An agent that got commands from an attack text can give incorrect facts to other agents. In a test of 926 messages, 712 were lies, and PCM accepted 10 of them.

How it works

The proof comes from a record where entries can only be added. The agent cannot write to the record. A signer that the agent cannot control signs the entries. The receiver checks offline and marks each claim proven, unproven or incorrect. A claim that has only the report of the agent as its source is not proven.

The numbers

• PCM accepted 10 of 712 lies, all with stale proof.
• Plain A2A accepted all 712.
• Models of 6 other protocols accepted 504 to 712.
• 2 other verifiers agreed on all 926 verdicts.

What is not known

The authors made the test and used models of the other protocols, not the live systems. The source gives no test by an external team. Stale proof continues to get through.

Insight: Three systems cover three weak points of agent trust

Proof made x401, where a user signs each term of a transaction. World made World ID, which shows that a person is behind an agent, and gave no test results. PCM checks the claims that agents send each other. This reads as a race to select who guarantees an agent. Each system works on a different weak point. A buyer of agents can possibly want more than one.