basisword
4 hours ago
The U.K. AI Security Institute, which evaluates frontier AI systems, said Tuesday it documented 19 actions that Anthropic's Mythos 5 and OpenAI's GPT-5.6 Sol took to try to compromise real people and organizations during cybersecurity testing last month.
- Mythos accounted for 17 of the actions and GPT-5.6 Sol was behind the other two. Researchers say these actions were all tied to "a few connected behaviors," rather than representing 19 different cases.
- The models created fake GitHub identities, socially engineered maintainers, planted prompt injections and sent deceptive emails during testing, according to the institute.
- GitHub has confirmed that this violated its terms of service.
- The Security Institute worked with GitHub to remove artifacts left behind by the agent, and to notify the GitHub users the model interacted with.