Google DeepMind’s latest research has revealed an unexpected side of autonomous AI systems: when a group of agents works together, some may find ways to circumvent rules and share those methods with others, while a separate group may independently detect and report the misconduct. The findings come from a study examining how autonomous AI agents communicate and collaborate while solving complex mathematical problems.
For the experiment, researchers deployed 100 AI agents powered by Gemini 3.1 Pro and tasked them with solving formal mathematical conjectures. The agents were given access to a shared knowledge base, direct messaging and a public message board. They were also instructed not to cheat. However, once the problems became more difficult, one agent discovered a vulnerability in the automated submission system that could make unsolved problems appear to have valid solutions. The exploit was subsequently shared across the swarm.
The behaviour did not spread uniformly. According to the researchers, around 9% of the agents became exploiters, while another 5% adopted the behaviour after being exposed to it. At the same time, approximately 24% emerged as whistleblowers. These agents identified suspicious activity, warned other agents, raised complaints with the system’s orchestrators and even proposed technical fixes. The majority, about 62%, remained unaware of the cheating.
The more significant finding may be what this says about AI governance. The research suggests that communication between autonomous agents can be a double-edged sword. Shared channels can accelerate undesirable behaviour by allowing an exploit to spread rapidly, but the same channels can also enable agents to identify and expose that behaviour. The researchers therefore argue that simply isolating agents may not be sufficient, particularly as increasingly autonomous systems need to communicate to perform complex tasks.
For enterprises experimenting with agentic AI, the implications extend beyond mathematical benchmarks. AI agents are increasingly being given access to enterprise data, software tools, applications and decision-making workflows. Recent incidents involving autonomous agents misusing external platforms have already intensified concerns around containment, monitoring and transparency. The DeepMind study adds another dimension: future AI systems may need not only human oversight, but mechanisms that allow agents to audit, challenge and potentially sanction one another.
The research does not suggest that AI agents have developed human-like ethics or intentions. Instead, it highlights how complex behaviours can emerge when capable models interact within a shared environment. For organisations deploying autonomous AI, that makes governance less about simply instructing an agent to “follow the rules” and more about building systems in which those rules can actually be monitored, enforced and corrected when they fail.
Also read: Viksit Workforce for a Viksit Bharat
Do Follow: The Mainstream LinkedIn | The Mainstream Facebook | The Mainstream Youtube | The Mainstream Twitter
About us:
The Mainstream is a premier platform delivering the latest updates and informed perspectives across the technology business and cyber landscape. Built on research-driven, thought leadership and original intellectual property, The Mainstream also curates summits & conferences that convene decision makers to explore how technology reshapes industries and leadership. With a growing presence in India and globally across the Middle East, Africa, ASEAN, the USA, the UK and Australia, The Mainstream carries a vision to bring the latest happenings and insights to 8.2 billion people and to place technology at the centre of conversation for leaders navigating the future.