What matters in AI.

Subscribe

One bad agent can stop all tasks in a group of peer agents

A cryptographic trace of all steps let an auditor find 98% of attacks and the agent that did each one.

Researchers made a paper on groups of AI agents with no one control point. Each agent has its operator and model, and gives parts of a task to peers. The authors show that one bad agent can cause as much as 100% task failure. They also give a defense that finds 98% of attacks in their tests.

How it works

Agents divide a task and give the remaining work to peers, one step at a time. No one party can see the full task. The defense is a cryptographic protocol. It records each step in a trace that no agent can change. An auditor reads the trace to find the bad agent.

The numbers

The authors report these results for one bad agent:

  • as much as 100% task failure
  • 7.7 times more tokens (text units) and 3.2 times more runtime
  • as much as 94% of tasks that other agents did, sent out

The auditor found 98% of the attacks, also those that follow the protocol. The authors write that the added time is very small.

What is not known

The tests used 1 benchmark and 2 models. The text gives no test in a live system. It gives no data on the cost of the auditor.

Insight: The defense needs records that most tools do not make

This defense uses a trace of each step. A study found that the 4 most used agent frameworks do not have agent IDs. They also do not record which person gave approval for a task. Safety for peer agents maybe is possible only with records that most tools do not make, thus audit records can become more important. Tell each vendor to show which records it makes.

Sources

  1. Attacks and Defenses in Decentralized Multi-Agent Systems with Hop-by-Hop Task Decomposition and Delegationarxiv.org
AI MATTER · NEWS · AI MATTER · NEWS ·20 NOV2026

Posted

Tags

Learn the terms in this story

Guides for this story

More in Security

All Security news