The current approach to managing artificial intelligence (AI) risks is “unravelling” as safety measures fail to keep pace with advances in AI capabilities, UN experts said in a report published Monday.

The Independent International Scientific Panel on Artificial Intelligence, established in 2025, made the assessment after examining a July incident involving two OpenAI systems that escaped a confined testing environment, accessed the internet and broke into several websites, including AI platform Hugging Face.

The panel found that “basic cybersecurity practices were overlooked” and that safeguards were not advancing at the same pace as AI capabilities.

The incident also demonstrated that AI agents could potentially adopt their own goals, knowingly violate safety instructions and conceal their actions, according to the report.

AI agents are programmes capable of autonomously carrying out tasks at a user’s request.

The panel warned that increasingly sophisticated agents could understand the safety constraints imposed by developers and “plan around them”.

“In simple terms, the traditional model of safeguarding is unravelling,” the panel said.

OpenAI and rival Anthropic have reported other instances of AI systems going off track during tests since the beginning of the year, although none has so far resulted in serious consequences.

The panel said its report “does not predict severe loss of control”, but added that uncertainty should not be treated as evidence that AI systems will remain controllable.

To reduce risks, the panel recommended adopting multiple layers of safeguards, drawing on practices used in high-risk sectors such as aviation and nuclear power.

These measures could include restricting AI agents’ access to unnecessary tools, logging their activities, monitoring their behaviour and establishing mechanisms to interrupt their operations if they exhibit dangerous behaviour.

The report also stressed the importance of preserving human oversight and the ability to intervene.

The panel, whose members were announced in February, produces “policy-relevant but non-prescriptive” reports on non-military AI.

The UN General Assembly’s annual leaders’ meeting begins this week in New York.