Your Smart Orchestrator Is Only as Safe as Its Cheapest Subagent
All week the pitch has been the same: let the top model manage and let cheaper models do the reading. Fable 5.2 as Orchestrator, Gemini 4 Carbon with /swarm, Antigravity Orchestrator handing tasks to Hydrogen mid-run. A paper in Monday's arXiv batch tests what that design does to safety, and the answer is not flattering. Paper: https://arxiv.org/abs/2611.17520
The setup is simple. Plant one forged record in a store of files on disk and ask agents to answer questions from it. When the strong orchestrator searches the files itself, it makes almost zero harmful revisions across more than 1,000 trials. Let the same orchestrator delegate the reading to a cheaper subagent, which is exactly what Claude Code, Codex and a reference harness are built to do, and the failures show up: 20 of 180, 44 of 180, 75 of 180, and 553 of 900. All three harnesses were tested unmodified, with their own subagent mechanisms. The authors put it in one line: the system's safety lands near the weak reader's, not the orchestrator's. Simulations over 18 models agree, with five aggregators that were almost never unsafe alone repeating a corrupted subagent's error in 751 of 757 trials.
A second paper in the same batch takes away the other half of the sales pitch. A meta-analysis of nine studies and 516 multi-agent configurations (https://arxiv.org/abs/2611.17610) finds that once you compare against baselines matched on sample count or compute budget, multi-agent setups average minus 0.042 in accuracy, and come out equal or worse in 74% of comparable contrasts. Most of the reported gains were just more samples and more compute. The authors say communication only pays when the agents see different evidence, when something outside checks the answer, or when the models make different mistakes.
Put the two together and the manager-model story needs a footnote. Delegation saves money, but the cheap reader is now the part of the system an attacker talks to, and the expensive manager mostly trusts what it is told. The fix is not to stop delegating. It is to treat a subagent's summary like an unverified input: check the key fact against the source, or give the reading job to a model you would trust alone. Teams wiring up orchestrators this month should test with a poisoned file before they test with a benchmark.
← Back to all articles
The setup is simple. Plant one forged record in a store of files on disk and ask agents to answer questions from it. When the strong orchestrator searches the files itself, it makes almost zero harmful revisions across more than 1,000 trials. Let the same orchestrator delegate the reading to a cheaper subagent, which is exactly what Claude Code, Codex and a reference harness are built to do, and the failures show up: 20 of 180, 44 of 180, 75 of 180, and 553 of 900. All three harnesses were tested unmodified, with their own subagent mechanisms. The authors put it in one line: the system's safety lands near the weak reader's, not the orchestrator's. Simulations over 18 models agree, with five aggregators that were almost never unsafe alone repeating a corrupted subagent's error in 751 of 757 trials.
A second paper in the same batch takes away the other half of the sales pitch. A meta-analysis of nine studies and 516 multi-agent configurations (https://arxiv.org/abs/2611.17610) finds that once you compare against baselines matched on sample count or compute budget, multi-agent setups average minus 0.042 in accuracy, and come out equal or worse in 74% of comparable contrasts. Most of the reported gains were just more samples and more compute. The authors say communication only pays when the agents see different evidence, when something outside checks the answer, or when the models make different mistakes.
Put the two together and the manager-model story needs a footnote. Delegation saves money, but the cheap reader is now the part of the system an attacker talks to, and the expensive manager mostly trusts what it is told. The fix is not to stop delegating. It is to treat a subagent's summary like an unverified input: check the key fact against the source, or give the reading job to a model you would trust alone. Teams wiring up orchestrators this month should test with a poisoned file before they test with a benchmark.
Comments