A multi-agent system is only as safe as its weak agent
When a strong model gave work to a cheaper sub-agent, the system was not safe in 553 of 900 trials in one test.
Researchers tell that Claude Code, Codex and a test harness can be broken by attacks that did not work on one model. A harness is the software that operates the agents. The team put an incorrect record in a set of files. A strong model that read the files with no sub-agent made almost zero dangerous changes in more than 1,000 trials.
How it works
The strong model, the orchestrator, gives work to a cheaper sub-agent. These tools are made to do this. The sub-agent reads the incorrect record. Thus the system has the safety of the weak sub-agent, not of the strong model.
The numbers
In one test, the same models and facts were not safe in 553 of 900 trials. In a simulation, aggregator models gave the error of a sub-agent in 751 of 757 trials.
What is not known
The paper tests one type of attack, an incorrect record. It does not tell how to stop it. It does not tell how many systems in use have this risk.
Insight: A strong model does not make a safe system
On 29 October 2026, a study found that the 4 most used agent frameworks do not always record which agent did a task. A paper of 12 November 2026 tells that one safe model does not make a safe group. A weak sub-agent can set the safety of the full system, and the records to find it are missing. Test the full system.
Sources
Posted