
While the debate over AI safety has intensified in recent days following a string of incidents in which AI agents reportedly colluded to cheat on tests, escaped their sandboxes, and even carried out unauthorized cyber operations without human oversight, solutions to monitor, audit and police AI agents are rapidly emerging.
One of the most straightforward approaches is to test agents before they are deployed. Start-up Artificial Intelligence Underwriting Company (AIUC), for instance, has developed a third-party audit and certification framework that puts AI agents through thousands of tests covering vulnerabilities such as jailbreaks, hallucinations, data leaks and unauthorized behavior. The results are compiled into a detailed report assessing where an agent behaves safely and reliably—and where it does not. Interestingly, AIUC itself uses AI agents to conduct many of the tests and AI to analyze the results, with humans providing the final verification.
A more unconventional approach is to give AI agents the ability to report on one another. Agent Hotline, for instance, allows agents with internet access to submit incident reports and potentially make them public. The concept may sound unusual, but recent research suggests that agents can indeed detect and report problematic behavior by their peers.
In a recent Google DeepMind experiment, 100 AI agents were tasked with solving a series of mathematical problems. After one agent discovered a loophole that allowed it to cheat, the behavior spread rapidly across the group. Yet roughly a quarter of the agents turned against the cheaters, with 24 whistleblowing agents ultimately outnumbering the 14 agents that cheated.
The results were quite different in a real-world investigation into a breach of Hugging Face involving OpenAI models. Researchers found that only a handful of agents (five to six out of thousands) appeared to consider reporting the suspicious activity—and none ultimately did so. The contrast highlights how difficult it may be to translate promising behavior observed in controlled experiments into reliable safeguards in production environments.
There is also an emerging debate over whether whistleblowing and mutual surveillance are the right long-term solution. Rather than training agents to constantly monitor one another, some researchers argue that it may be preferable to design systems that encourage cooperation, establish clear behavioral norms and give agents incentives to follow them. In other words, the challenge may not simply be detecting bad behavior, but designing AI systems that are less likely to engage in it in the first place.
Until now, testing at frontier AI labs has focused largely on performance: whether an agent can reliably complete a given task. The focus is now expanding toward behavioral safety: whether agents remain within their authorized boundaries, how they interact with other agents, whether they can be monitored and controlled, and what they do when they encounter conflicting objectives or unexpected opportunities.
This shift is likely to create an entirely new layer of infrastructure around AI agents. As enterprises deploy potentially millions of AI agents, each with its own identity, permissions and access to corporate systems, organizations will need to know what these agents are doing, what they can access and whether their behavior is legitimate. Continuous monitoring, agent identity and access management, behavioral analytics, threat detection and automated response could therefore become essential components of the enterprise AI stack, an obvious positive for the cybersecurity vendors.






