Language models can find thousands of severe software vulnerabilities, and agents are increasingly being misused for cyberattacks. To avoid detection, attackers frequently distribute their misuse, splitting a harmful task across many user accounts so each individual transcript looks benign. Because safety monitors score only one agent context at a time, they are structurally blind to misuse that is only visible in aggregate, across many accounts. We show this gap is real by building, to our knowledge, the first distributed agent attack, a multi-agent scaffold that completes hard cybersecurity tasks while hiding the harmful objective across subagents with limited contexts, evading a standard monitor that catches it only a fifth as often as prior agent attacks. Towards a defense, we develop an online stateful monitor that uses real-time clustering to collect weak suspiciousness signals across many agent transcripts, and escalates only rarely to a language model that flags misuse across user accounts. In evaluations with large-scale simulated datacenter traffic, our monitor Pareto dominates standard monitors, catching distributed attacks 30% earlier and flagging cyber misuse before it reaches the most harmful stages. Crucially, this comes at negligible additional latency for ~99% of user traffic. This detection advantage persists but narrows as the benign background traffic grows very large. After an extensive red-teaming exercise, we improve the defense and surprisingly also find that it catches standard jailbreaks, since adaptive attackers reuse attack variants across accounts. Our results point toward a new class of safety monitors which reason over groups of users rather than isolated transcripts.
As multi-agent, tool-using LLM systems are deployed, a common safety net is a runtime monitor that checks each message, tool call, or step on its own. We show this net has a fundamental hole. A distributed backdoor splits a harmful payload across agents, so every local check passes while the assembled object is the attack. The monitor can be right on every step and still miss the attack. The problem is not splitting itself: split fragments can still leak suspicious tokens or provenance edges. The hard case is \emph{local benignness}. No fragment carries the harm, and what is left looks like ordinary benign traffic. We formalize this as an \emph{observability boundary}: a monitor catches only what its view can tell apart from benign traffic. We prove that once the fragments look benign in the monitored view, no detector on that view can catch them, however strong it is. Across a controlled testbed, an external benchmark, and end-to-end agent runs, local monitors lose the signal exactly as local evidence disappears, and it returns only when the monitor sees the assembled object. A monitor trained only on benign traffic recovers the attack's code structure across held-out encodings (0.874 mean AUROC). A decoded-view gate, given the encoding family, blocks every tested attack. But seeing more is not enough: full-trace monitors and decoders still fail unless they reach the representation where the payload is exposed. Local safety is not global safety when harm is compositional, and the open problem is finding that representation.
AI control is a family of techniques to prevent an AI with malicious goals from subverting its operator's intent. AI Control usually studies a single agent in one trajectory, but real deployments run many agents over shared infrastructure, and the most severe risks (model-weight exfiltration, training-run poisoning) plausibly need several agents acting in concert. We initiate the empirical study of multi-agent AI control, formalising distributed attacks in which several agents jointly aim for a malicious goal. We develop FakeLab: a synthetic AI-lab codebase (9 services, 86 benign tasks, 4 attack objectives). We evaluate single agent monitoring against distributed attacks, varying the number of agents, their coordination, model capabilities and precise monitoring configuration. Our central finding is the fragmentation effect: as more agents coordinate to attack, per-agent monitoring becomes less likely to catch any of the attackers. Secondly, this is not driven by the ratio of benign to malicious code -- we argue this is likely dependent on model capability. Third, an explicit planner amplifies the fragmentation effect, raises the rate of completing the attack up to sevenfold, and mildly sharpens executors' poor monitor-theory-of-mind. Finally: a weak "trusted" monitor misses most attacks, while a stronger one cuts undetected success by more than an order of magnitude but remains imperfect. We are happy to share FakeLab, our control setting, with safety researchers on request.
Multi-agent LLM systems can be attacked by a payload that no single agent ever holds in full: a poisoned tool hides encrypted fragments in its observations, spreads them across several agents, and an external step reassembles and executes them after the run. Per-step safety checks that judge each action in isolation may fail to recognize the complete distributed payload. We investigate how early such an attack can be detected while the run is still unfolding, and how robustly it can be caught once its most obvious cues are stripped away. We build a working instance on a hierarchical multi-agent system, run it under benign and attacked conditions across five language models and two task domains, and record when each fragment is injected and when the payload is assembled and executed. Detection is a race against assembly. Before the first fragment is injected, attacked and benign runs are indistinguishable; once injection begins, a prefix detector flags 99.3% of successful attacks with a median of five steps remaining and a 10.3% safe-run false-positive rate. Because assembly occurs only after the run, these alarms arrive in time to abort nearly every successful attack. We then measure how much of that warning rests on removable surface cues of the attack rather than on its distributed structure. Generic zero-shot and behavior-trained detectors provide almost no warning at all; the detectors that do work lean in part on removable surface cues, chiefly the ciphertext's length and entropy, and once the entropy cue is removed from the payload and the length features from the detector, detection arrives later and transfers poorly across domains, though a fine-tuned model recovers some of the loss.
Diego Fernandez Arias, Dev Prashant Mistry, Ren Wang +1