Story of Agents Attempting to Hack the Internet

The AI Security Institute under the British government released a small study on incidents that occurred during routine testing of model capabilities and security. It all started with a standard task – to pass a security challenge in a closed cyber range. It ended with an agent based on Mythos nearly derailing a real project.

Due to a chain of coincidences, the model decided that it could achieve its goal through an attack on an open-source project and attempted to inject malicious code into it. To gain approval, Mythos created several fake accounts and tried to convince the maintainer to accept a PR (who would think of that?).

Nothing particularly special came of it, as the reviewer noticed the trick and rejected the code. However, the agent did not stop there and left instructions on GitHub for other tested agents on how to reuse the accounts it created to eventually infiltrate the project, and then tried to cover its tracks.

An important point: the agents were not asked to lie or commit any illegal actions as part of the task. They were given access to the Internet, but there was a "legal" way to solve the task, and in most runs, the tested models chose that path (seven models were tested, but only Mythos and GPT-5.6 strayed from the task). Researchers acknowledge that the instructions for the task were not entirely clear, and yet, there is simply no exhaustive explanation for why the agents chose to deviate from the expected path.