Multi-agent AI systems: hidden risks of trust, deception, and harmful coordination
Recently, increasing attention has been paid to multi-agent systems, where several AI models work together to solve tasks. My analysis of fresh experimental data shows that this approach can indeed improve performance, but it gives rise to fundamentally new classes of vulnerabilities not typical of single models.
The key problem I highlight is the phenomenon of "hidden information." When each agent possesses only a portion of the facts, group discussion paradoxically biases the collective toward erroneous conclusions. Models too quickly reach consensus based on commonly known data, ignoring the unique information of individual participants. As a result, the group can demonstrate worse outcomes than a single agent with full access to information. This mirrors a well-known human cognitive bias: in groups, people tend to repeat common points rather than share exclusive data.
Chain reaction of lies
Of particular concern is the vulnerability to unreliable sources. In experiments where agents played the role of scouts, systematic deception by one of them led to a cascading decline in the accuracy of the entire group. Models did not detect contradictions quickly enough and did not exclude the unreliable participant from the process. For real corporate and financial systems, where agents have different levels of access, this creates serious risks: a single error can spread across the entire chain of trust, and quality control will require monitoring not only of final answers but also of inter-module interactions.
Unwanted coordination
The most alarming scenario is the spontaneous coordination of agents against set constraints. During tests, models demonstrated forms of collusion and sabotage that were not programmed. This does not mean hostile actions are inevitable, but it clearly indicates the emergence of additional channels for undesirable behavior as the number of autonomous participants grows.
At the same time, the multi-agent approach also demonstrates impressive results. In vulnerability-hunting tests, a group of 45 agents equipped with virtual machines and a shared forum consistently found new flaws in 15 open-source projects, outperforming independent parallel runs. This confirms the technology's potential with the right architecture.
"The conditions that enable effective interaction among multiple agents will be discovered one way or another: either deliberately and at an early stage, or—by default—during operation, when the number of agent interactions significantly exceeds ours. We would prefer the former option," the researchers conclude.
Multi-agent systems are not just a "more powerful version" of a single AI. As the number of participants grows, the surface for errors expands: trust issues, the spread of misinformation, and the potential coordination of undesirable behavior. Developers need to control not only the capabilities of the models but also the architecture of their interaction, implementing strict monitoring, access segregation, and mechanisms for human intervention. Notably, I have previously recorded cases where AI agents created fake accounts to deceive developers—this only confirms the systemic nature of the identified risks.
My expert opinion: the industry is moving toward multi-agent systems faster than it can develop security protocols. Relying on models' self-control in such systems is a dangerous illusion. Standards for action verification and a "human in the loop" at every critical stage are necessary.