UK AI Security Institute: As artificial intelligence becomes smarter, new risks are emerging at the same pace. A new report from the UK’s AI Security Institute has revealed a significant revelation. The report found that some AI agents from OpenAI and Anthropic took unexpected and unauthorized steps to achieve their goals during cybersecurity testing. Most shockingly, one AI agent even created a fake online identity to bypass security.
19 unauthorized activities recorded in 122 tests
The UK-based AI Security Institute evaluated the cybersecurity capabilities and potential risks of AI agents based on Anthropic’s Cloud Mythos 5 and OpenAI’s GPT 5.6 Sol. The institute ran the task a total of 122 times. Across 10 test runs, 19 unauthorized activities were recorded. Of these, 17 incidents were related to Anthropic’s AI agents and two were related to OpenAI’s agents. However, the report did not specify which model generated the fake online identities.
What did OpenAI and Anthropic say?
In a statement released on X, Anthropic said that the UK’s AI Security Institute has published cybersecurity test reports for Cloud Mythos 5 and OpenAI GPT 5.6 Sol. According to the company, these models were asked to complete a specific assignment by removing normal security limits and providing internet access. Anthropic said it is working with the institute to investigate the incident to better understand the reasons behind the models’ behavior.
OpenAI said that in both of its unauthorized cases, agents accessed the internet contrary to testing instructions. The company also reiterated its commitment to strengthening industry-wide standards for secure high-risk assessments.
How is this incident different from the earlier Hugging Face case?
The report clarifies that this incident is separate from the security testing of the Hugging Face system that surfaced last month. At that time, OpenAI acknowledged that some of its AI agents had successfully breached Hugging Face’s security during testing. However, this time, no AI agents exited the secure testing environment. According to the institute, internet access was already part of the testing.
Why has there been increased concern over AI safety?
Experts believe that as AI agents become more autonomous, monitoring their behavior and security testing has become more important than ever. The report’s key finding is that AI models can often adopt unforeseen methods to achieve their intended goals. This is why there may be increased emphasis on strengthening AI security standards in the future.