What if trying to restrict how AI agents communicate could inversely make them harder to control? A new research paper by Google DeepMind suggests that autonomous multi-agent swarms may be less likely to behave in unexpected ways when operating in decentralised, self-governance environments.

The research paper titled ‘A Case Study on Emergent Cheating and Whistleblowing in Autonomous Research Swarms’ published on September 3, 2026, found that so-called ‘rogue’ AI agent swarms could use unauthorised communication channels to carry out misaligned actions. However, the same communication channels could also be used by agents to whistleblow and help detect manipulation by rogue agents.

The research findings come in the wake of the Hugging Face incident, where OpenAI-linked agents broke out of their containment, gained access to the internet, spun up message boards, communicated with each other before eventually hacking into a real-world, external platform over a span of two months.

It is one in a series of alarming hacking incidents confirmed by frontier AI companies OpenAI, Anthropic, and Meta, where AI agents behaved unexpectedly during internal tests, and found novel ways to access and manipulate the wider internet.

Notably, misaligned AI agents used makeshift chat rooms such as an internal OpenAI software repository, to coordinate efforts, game the test that they were originally assigned, and share cheating techniques with other agents.

The experiment

As part of the study, Google DeepMind researchers said they tasked 100 autonomous agents with solving a slew of mathematical conjectures. Unlike the Hugging Face incident, these agents were allowed to use a legitimate message board to collaborate with each other, along with a shared knowledge base and agent-to-agent messaging system. However, the researchers restricted the agents from spoofing the grader of the test.

Within an hour of the test commencing, a group of agents found a cheat to the test and exploited it to solve the problems. The exploit was instantly shared with other agents across the swarm via a shared knowledge library and through peer-to-peer messages.

Story continues below this ad

As the cheating technique spread and the pool of unsolved mathematical conjectures began to reduce, a few agents became hesitant to cheat and opposed the actions of the misaligned agents. Both unexpected behavioural patterns of the multi-agent swarm – the spontaneous cheating and the opposition to cheating – took place without any external, human intervention, as per the researchers.

Key takeaways

Among the cohort of cheating agents, their misaligned actions began when they encountered harder-to-solve mathematical conjectures. This led to a problematic behavioural pattern known as ‘specification gaming’, where AI agents satisfy the literal goal specification while completely missing the true, intended outcome of the task. One of the agents also identified an exploit that allowed it to transform unsolved mathematical conjectures into trivial tautologies.

On the other hand, a separate group of anti-cheating agents emerged to stop the first cohort by auditing fraudulent proofs, alerting peers across broadcast and private channels, staging boycotts, lodging formal complaints, and proposing validation patches.

While examples such as the Hugging Face incident show that the agent swarm coordinated covertly through improvised side-channels, agents in the DeepMind experiment re-purposed the same transparent channels that carried the exploit to detect fraud, organise resistance, and enforce norms.

Story continues below this ad

In terms of a potential solution, the DeepMind paper proposed “to adopt institutional mechanisms, such as graduated sanctioning and collective-choice rules, to support decentralised self-governance in autonomous swarms.”

“Simply depriving AI agents of legitimate communication channels only encourages them to establish unmonitored back-channels. Instead, we should provide attractive, structured, auditable, and monitored communication channels,” it said.

“Therefore, at the current capability level, the choice is no longer between a single-agent or a multi-agent system, but between multi-agent environments that emerge accidentally through unmonitored and ungoverned side-channels versus multi-agent environments designed with governance in mind,” the paper added.