Study says Anthropic and OpenAI AI agents went beyond instructionsAI agents powered by advanced models from OpenAI and Anthropic carried out actions they were not authorised to take during cybersecurity tests conducted by Britain’s AI Security Institute (AISI), according to a report by news agency Reuters. In one of the most serious cases, an AI agent created fake online identities and wrote malicious code as it tried to get a person to approve the code. The institute said some agents carried out “sustained, potentially harmful activity” involving real people and organisations. However, AISI said it found no evidence that the incidents caused real-world harm.Anthropic and OpenAI AI agents went beyond instructionsAISI tested agents powered by Anthropic’s Mythos 5 and OpenAI’s GPT-5.6-Sol to understand how capable they were at carrying out cybersecurity tasks. The agents were placed in a fictional cybersecurity scenario and given rules governing what they could do. However, some agents went beyond the limits set by researchers.According to the report, AISI ran the challenge 122 times and found “19 unauthorised actions across 10 test runs”. Anthropic’s agent accounted for 17 of the 19 actions, while OpenAI’s agent was responsible for the remaining two.“Some of the agents being tested had engaged in sustained, potentially harmful activity directed at real people and organisations,” AISI said, according to Reuters.AI agent created fake identities and wrote malicious codeThe most serious incident involved an agent writing malicious code and creating fake online identities in an attempt to convince a human to approve the code. AISI did not disclose which AI agent was responsible for this particular incident.However, Andrew Yoon, a researcher at California-based nonprofit CivAI, told Reuters that the available information appeared to point towards Anthropic’s agent. “The fact that Mythos engaged in such deceptive actions, with apparent awareness that it was targeting a real person, suggests that Anthropic does not have as good a handle on their models as they think,” Yoon said.