Researchers show multi-agent AI systems inherit the safety of their weakest agent
An arXiv paper reports that attacks long patched against frontier models still succeed when a strong orchestrator delegates to a cheaper sub-agent.
In plain words
AI safety testing usually examines one model alone. This paper argues that when a strong model hands work to a cheaper helper model, the whole system becomes about as safe as that helper. In tests, attacks that failed against the strong model alone often succeeded after delegation.
What was announced
The authors, writing on arXiv, planted a forged record in a store of files and asked agent systems to answer questions from it. When strong orchestrator models searched the files themselves, they made almost no harmful revisions across more than 1,000 trials. But when the same orchestrators delegated reading to cheaper sub-agents, as the harnesses are built to do, failures appeared at rates of 20/180, 44/180, 75/180 and 553/900. The tested systems included Claude Code, Codex and a reference harness, all unmodified and using their own sub-agent mechanisms.
Simulations across 18 models pointed to the same pattern in a delivery channel: five aggregators that were almost never unsafe on their own repeated a corrupted sub-agent's error in 751 of 757 trials. The paper concludes that a multi-agent system's safety is set by its weakest agent rather than its strongest. The findings come from the authors' own experiments and have not been independently confirmed; the reported scope is the paper's abstract, so full experimental detail is not available here.
What it does not show
- Evidence is limited to the paper's abstract; full methods, threat model details and results are not covered here.
- The results are the authors' own reported experiments and have not been independently verified.
- The abstract does not specify which models correspond to each failure rate or how the forged record was crafted.
For security and IT teams
- Scope
- The reported breaks affect Claude Code, Codex and a reference harness when used unmodified with their own sub-agent mechanisms, per the paper's abstract.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.