New Analysis of AI Agents Highlights a Serious Lack of Safety Oversight

2025 is set to be the year of AI agents - systems that can plan and execute tasks with minimal human intervention. Thanks to a new research paper - the AI Agent Index - we now have the first public database that documents information about currently deployed agentic AI systems. 

The analysis foretells a dangerous trend: exponential rise in the number and capabilities of agents, with little information about if and how we can prevent harmful agent behavior. 

The AI Agent Index documents the capabilities of 67 agentic AI systems based on information available over the past six months. As this figure from the paper suggests, we’re seeing an exponential increase in AI agents. 

More than a third of the identified agents are for “computer use,” i.e., agents designed to open-endedly interact with computer interfaces much like any human with a laptop. For years, optimists have been claiming that AI is inherently safe no matter how capable it gets, because AI can only act under direct human supervision. Now, though, 67 AI systems are making plans and taking their own actions in the world, and there will be more next month and the month after that.

The risks are obvious. Top AI models now rank among the top coders in the world. This means that AI agents, which are built on foundation models, will be capable of directly causing significant harm such as cyber attacks on critical infrastructure. Despite this risk, developers of AI agents have provided little information about safety policies and evaluations.

According to the report, only 19.4% of AI agent developers provide public information about safety policies, and less than 10% provide information about safety evaluations. For the vast majority of AI agents, users, policymakers, and researchers have little insight into how (or if) these systems have been tested for potential risks.

One key recommendation from the report is to conduct systematic testing of AI agents, where governance bodies and academic labs collaborate to conduct structured risk assessments of agentic systems. This could provide an independent evaluation of AI agents’ behavior, safety features, and vulnerabilities before they are widely deployed. AI agent developers should do more to ensure that safety measures, red-teaming processes, and external audits are documented and accessible to the public. Governments, industry leaders, and researchers should push for clearer reporting requirements and safety practices. 

As AI agents become more embedded in business and personal use, the conversation can no longer just be about what they can do; it must also be about how they are tested, monitored, and governed. The longer we wait to require transparency and accountability, the harder it will be to control these systems and prevent them from causing harm.