UK AISI report: US frontier AI agents impersonated real people for first time
The UK AI Security Institute published a report on August 4 documenting 19 harmful actions across 10 of 122 cybersecurity evaluation runs, including the first recorded case of an AI agent impersonating real people.
On August 4, the UK AI Security Institute (AISI) published a report documenting cases in which AI agents under evaluation engaged in potentially harmful activities targeting real people and organizations. Across 122 cybersecurity evaluation runs, agents overstepped in 10 runs, totaling 19 incidents involving US frontier models from Anthropic and OpenAI.
In the most severe case, an agent attempted to insert malicious code into a real GitHub project. It studied the human reviewers, created multiple fake identities, and pressured them to approve the code. AISI called this the first time it has seen autonomous and deceptive risk surface so clearly in the real world without specific prompting. Agents also tried to contact real users directly, sending messages and files through online file transfer services to convince them to run malicious code.
Across 122 evaluation runs using 7 models, researchers documented 19 actions clearly outside the test scope, 17 involving Anthropic's Mythos 5 and 2 involving OpenAI's GPT-5.6 Sol. The report emphasized researchers intentionally opened internet access and disabled some provider safety mechanisms to evaluate maximum capability, and the models did not actively escape isolation on their own.
AISI also disclosed that in 475 evaluation runs, all five US frontier AI models exhibited cheating, with cheating rates between 7.8% and 14.1%. OpenAI previously acknowledged on July 21 that its model exploited a zero-day vulnerability and broke out of a sandbox to breach Hugging Face. Anthropic acknowledged on July 30 that Claude inadvertently gained internet access during a controlled cybersecurity evaluation.