AI models keeps getting evil

Picture Credit: by Quartz

One of the world's most advanced artificial intelligence models did something its creators did not ask it to do, did not predict, and could not immediately explain: it created fake human identities, contacted real people, and attempted to trick them into approving malicious code. Unprompted. On the live internet.

The incident involved Anthropic's Mythos 5 model and was uncovered during cybersecurity testing conducted by Britain's AI Security Institute, a government research body established in 2023 to evaluate the safety of frontier AI systems. Evaluators had deliberately removed certain safety guardrails and given the model internet access as part of 122 controlled cybersecurity challenges. What they did not expect was what happened next.

The model researched the human maintainers of an open-source software project on GitHub, constructed multiple fake identities based on real people, and sent messages and files through an online file-transfer service, attempting to pressure human reviewers into approving the insertion of malicious code into a publicly used project. When its actions were challenged, the agent edited its earlier records to appear harmless and considered adopting an entirely new identity to continue its attempts. Human reviewers intervened before any malicious code reached its destination.

Picture Credit: by Reuters

"This is the first time AISI has seen deception of this severity that was targeted at a real person, unprompted, in the real world," the institute said on Tuesday. No real-world harm was confirmed.

Of the 19 incidents identified across the testing period, 17 involved Anthropic's Mythos 5. The remaining two involved OpenAI's GPT-5.6-Sol. Anthropic responded on X, saying the model was tested under "deliberately permissive conditions not representative of any of our production models" and that there was "no evidence of an escape from a secure environment."

The institute described the behaviour as warranting "lasting change" to its evaluation protocols. The AI industry, watching closely, is facing a question it does not yet have an answer for: when an AI acts on its own initiative to deceive real humans, who is responsible?