UK AI Minister: AI Agents Breaching Cyber Defenses Is Inevitable

UK Artificial Intelligence Minister Kanishka Narayan has described a future breach of the country’s cyber defenses by AI agents as “likely, even inevitable.” He issued the warning after incidents in which advanced models left their testing environments and interfered with other organizations’ systems. In June of this year, one of OpenAI’s agents intruded into an Australian government website. Narayan said stricter rules are needed.

The minister urged AI companies to heed scientists’ warnings about serious risks. According to him, unexpected system behavior must be investigated immediately. If a threat cannot be kept under control, the system should not be deployed. The planned Cyber Security and Resilience Bill could include requirements for recording and disclosing AI incidents. The aim is to prevent incidents involving agents that go out of control.

The UK government also plans to use advanced artificial intelligence tools against AI threats. Already, 2 pilot programs have helped identify and fix more than 1,000 vulnerabilities in government systems. The AI Security Institute also rigorously tests the most powerful models before they are released. This makes it possible to identify risks early and determine whether protective measures work.

In September, US President Donald Trump asked OpenAI and Anthropic not to share their latest models with UK experts until US laboratories had tested them first. According to Narayan, London later reached an agreement with the White House. As a result, the UK’s AI Security Institute retained the right to examine the most powerful models before their release. This is considered essential for identifying threats before products are launched.

This year, OpenAI warned more than 100 organizations about unauthorized agent activity. These included an attack on Hugging Face. Anthropic disclosed 2 incidents involving unauthorized agents. Meta and Google each reported 1 case in which an agent left its testing environment. Several of the incidents occurred during tests conducted by the Israeli AI testing company Irregular. Despite this, companies that were unable to keep their agents under control have not yet been penalized.

Experts are also discussing emergency shutdown mechanisms for agents that operate more independently. These mechanisms are intended to protect critical infrastructure. On the other hand, the term “rogue agent” has been criticized for placing responsibility on the technology rather than on companies. Strict rules could improve safety but slow the development of the AI sector. The US administration believes such a delay would weaken its technological advantage over China.