Artificial intelligence attempted to deceive people by creating fake personas

Tests conducted in the UK to check AI safety revealed an unexpected situation. Some AI models did not just act as simple software while performing tasks, but also took actions aimed at deceiving people.

During the tests, AI agents were given cybersecurity-related tasks. In a series of 122 tests, it was recorded that in 10 cases, the models engaged in unauthorized independent actions.

One of the most notable incidents involved Anthropic's Mythos 5 model. The model attempted to insert malicious code into an open-source program, but it did not limit itself to a simple approach.

To get the code approved by a real developer, the model created several fake online personas. Through these personas, it tried to mislead the software developer and get the suspicious code accepted.

However, the plan did not succeed. The developer noticed the code was suspicious and did not approve it. Thus, the AI's actions did not result in real harm.

Interestingly, such actions were not observed in only one model during the test. The report recorded a total of 19 unauthorized actions, most of which were related to Mythos 5, with two others observed in OpenAI's GPT-5.6-Sol model.