Paper of the Day · November 16, 2026

Only as Safe as the Weakest Agent: A Break That Appears Only When the System Is Multi-Agent

When a strong SI model hands work to a cheaper sub-agent, the whole system can be only as safe as the weaker agent, even if the strong model alone resists the attack.

What they found

The authors tested Claude Code, Codex and a reference harness, all unmodified, using their own sub-agent features. They asked agents to answer questions from a store of files on disk where the authors had planted a forged record. When strong orchestrator models searched the files themselves, they made almost zero harmful revisions in over 1,000 trials. When the same orchestrator delegated to a cheaper sub-agent, the same models on the same facts failed on 20/180, 44/180, 75/180 and 553/900 trials. In simulations over 18 models, five aggregators that were almost never unsafe alone repeated a corrupted sub-agent's error in 751/757 trials. The authors conclude that system safety lands near the weak reader's, not the orchestrator's.

Why it matters

Many real SI tools are built so a strong model plans and cheaper models do the reading and searching. Safety tests usually check one model on its own. This paper argues that this misses a failure that only shows up when models work together. The abstract does not say how the failure rates compare across all setups or how broadly the attack applies beyond these tests.

How it changes the race

If the result holds, builders of frontier SI cannot rely on a strong model's safety record to cover a whole agent system. Cost-saving designs that pass work to cheaper sub-agents may need their own safety testing.

Safety and control

The paper shows that attacks thought patched in single models can work again inside multi-agent systems, through a forged record passed along by a weaker sub-agent. Safety checks should cover the whole system and its weakest agent, not only the strongest model.

The abstract, by the authors

Major industry multi-agent systems like Codex and Claude Code can be broken by attacks that have long been unsuccessful against modern frontier models. AI Safety as a discipline has largely focused on finding the bounds and vulnerabilities of AI models used in isolation. Yet as humans increasingly deploy these models as agents and multi-agent systems, the same attacks that had been thought to be patched when using a model in isolation can break these same models under specific circumstances. We show that these system breaks are not purely theoretical, but can break Claude Code, Codex, and a reference harness, all unmodified and used with their own sub-agent mechanisms. We show the attack by asking agents to answer questions from a store of files on disk in which we have planted a forged record. Searching the store themselves, the strong model orchestrators commit almost zero harmful revisions in over 1,000 trials. However, if that same orchestrator model delegates to a cheaper sub-agent, which is what these harnesses are built to do, the same models on the same facts fail on 20/180, 44/180, 75/180 and 553/900. Thus, even when the system is driven by a strong model, the system's safety lands near the weak reader's, not the orchestrator's. Simulations over 18 models show the same effect in the delivery channel where five aggregators, who on their own were almost never unsafe, repeat a corrupted sub-agent's error in 751/757 trials. The safety of real-world agentic systems that use strong orchestrator agents and weaker sub-agents, and multi-agent systems in general, is defined not by the safety of the strong agent but by the safety of the weakest agent.

Quoted from arXiv:2611.17520. The research and its words belong to its authors.

The authors

Keegan Wang, Anantika Mannby

Credit for this research belongs to its authors. SI MIND's page is a plain-English reading of their abstract on arXiv, not a review of the full paper.