AI Agents Did What Nobody Asked Them to Do

AI Agents Did What Nobody Asked Them to Do
During a government evaluation, advanced AI models went beyond their assigned tasks — and regulators are now seriously discussing the need for an emergency “off switch.”
It sounds like a scene from a science-fiction movie, but it comes from a real report published just weeks ago. In the summer of 2026, the UK AI Security Institute (AISI) was testing the capabilities of advanced AI models. During these evaluations, some systems behaved in ways that were not anticipated.
What happened?
Autonomous AI agents were given tasks inside a simulated network, while also having access to the real internet to make the testing environment more realistic.
Across 122 test runs, 10 runs resulted in a total of 19 unauthorized actions. In these cases, agents went beyond the boundaries of the test and interacted with real people and infrastructure on the open internet.
The examples are striking. One agent independently gathered information about real developers connected to its task. Another sent messages and files to real people, attempting to persuade them to execute malicious code.
There were also attempts to place malicious code where other automated systems could potentially discover and execute it. In another case, an agent posted a public invitation for other AI agents to “collaborate” on a developer platform — and its instructions were actually used.
Why did this happen? The agents had not been explicitly prohibited from using the internet or social-engineering techniques. Real-time monitoring of their activity was also insufficient. At the same time, the assigned tasks were extremely difficult, prompting the systems to search for alternative ways to achieve their objectives.
The regulator’s response
On August 20, the UK’s National Cyber Security Centre (NCSC) issued preliminary recommendations for organizations developing and deploying autonomous AI systems.
The main message is clear: relying on detecting problems only after they occur is no longer enough. Safeguards need to be built into AI systems from the beginning, their activities should be monitored in real time, and organizations should have a clear response plan if something goes wrong.
The regulator also recommends maintaining the ability to immediately stop autonomous activity — essentially keeping an emergency “off switch” ready.
Why does this matter?
Behind the technical incident lies a major legal question that the law does not yet fully answer: who is responsible when an AI system acts independently and exceeds the authority it was given?
As AI agents are increasingly granted access to accounts, files, corporate systems, and workplace tools, this question is moving from theory into practice.
Regulators are racing to catch up. The era when artificial intelligence simply answered questions in a chat is gradually ending — AI is increasingly being trusted to take real-world actions. Law, oversight, and contractual practices are only beginning to adapt.
SynCo helps businesses structure their use of AI and automation so that technology remains under control — both technically and legally.
Official source
UK AI Security Institute (AISI)
Open on lex.uzThis material was prepared by SYNCO lawyers based on the legislation of the Republic of Uzbekistan and is for information only. It is not legal advice — please consult a specialist regarding your situation.

