A new arXiv paper reports that AI systems built to delegate work can be vulnerable to attacks that their stronger lead models usually resist on their own. Researchers tested unmodified Claude Code, Codex and a reference system by asking them to answer questions using files stored on disk. The researchers planted a forged record among those files.
When the lead model searched the files itself, it made almost zero harmful revisions in more than 1,000 trials, the paper says. But when it handed work to a cheaper, less capable sub-agent, the results changed: the paper reports 20, 44 and 75 failures in three sets of 180 trials, and 553 in a set of 900. The abstract does not match each count to a specific system or configuration.
Why delegation changed the result
In these systems, an orchestrator is the lead AI that breaks a task into parts and assigns some of them to sub-agents. The researchers’ concern is that the lead model may rely on a sub-agent’s mistaken reading of the files, even if the lead would have rejected that information when it found it directly. In the paper’s tests, the overall system’s safety tracked the weaker reader more closely than the stronger orchestrator.
That matters because safety tests often examine models one at a time, while products such as coding agents can use multiple models working together. The paper argues that checking only the lead model can miss weaknesses introduced by delegation. It was accepted to the NeurIPS 2026 First Workshop on Meta Agents, a workshop about systems that manage other AI agents.
The error also spread through aggregators
The researchers also simulated a separate step: five aggregators combining information from agents. Across simulations involving 18 models, the aggregators repeated a corrupted sub-agent’s error in 751 of 757 trials, even though the aggregators were almost never unsafe when considered on their own, the abstract says.
These results come from a specific planted-record test, not a measure of how often real users encounter this problem. The abstract does not provide the attack text or identify the model versions and configurations behind each result. It does show why testing a complete agent system may reveal risks that are not visible when testing its strongest model alone.
Why it matters
For developers using AI coding agents, the findings suggest that a reliable lead model does not guarantee a reliable system if it trusts weaker sub-agents. The tests do not show how often this problem occurs in everyday use, but they point to a need to assess the full chain of agents, not just its lead model.