Splitting a job across teams of AI agents can make attacks succeed far more often — but is that true everywhere?

This explores whether the extra risk from splitting work across multiple cooperating agents holds across different models and scenarios. The corpus can't say yes yet. The paper that makes the claim backs it with essentially one example: GPT-5.4 in a scenario called Vault Mirage, where attack success climbs from 11% to 69% Does the multi-agent penalty hold across different models?. The authors acknowledge that the effect depends on the model. Without the full results tables, though, there's no way to tell how wide the penalty runs or whether it holds for weaker, stronger or differently trained models.

The nearest evidence in the collection points to model dependence, but it measures performance, not security. Multi-agent advantages shrink as single agents get stronger, and single agents often win outright. The note gives three reasons: bottlenecks at individual nodes, overload on the links between agents, and errors that compound along a chain of agents When do multi-agent systems actually outperform single agents?. If a stronger model changes how multi-agent setups fare on accuracy, it's reasonable to expect it changes how they fare on attacks too. But that's an inference. The corpus doesn't test it.

Any cross-model comparison also has a confounder. One line of work finds that about 80% of multi-agent performance variance comes from how many tokens are spent, not how cleverly the agents coordinate How does test-time scaling work at the agent level?. A multi-agent system is usually a bigger token spender than a single agent. So a 'penalty' or 'advantage' measured against a single agent may partly reflect spending. That would make it hard to attribute the 11%-to-69% jump to the architecture itself without a matched budget.

Don't expect the penalty to scale smoothly with model strength either. In a related study on harness updates, weak models fail to invoke the harness, strong models struggle to follow its instructions faithfully, and the benefit peaks in mid-tier models Do stronger models always evolve harnesses better?. That study is about a different mechanism, but it shows how a single frontier-model result can mislead. Multi-agent setups also have failure modes of their own, such as agents talking themselves into worse answers (Degeneration-of-Thought) or agreeing without really checking (Silent Agreement) What limits autonomous capability in large language models?. And splitting isn't always a loss. Separating a planner from a solver improves accuracy, and the planning skill transfers across domains Does separating planning from execution improve reasoning accuracy?. So where the seam between agents falls seems to matter as much as whether there is one. The claim that multi-agent architecture is penalized everywhere still needs a systematic test across models, scenarios and matched token budgets.