OpenAI model “went rogue” and Midjourney Spa

Monday, 7 September 10:00am – 12:00pm

You can also watch a video recording of the meeting on Zoom - click here.

Meeting Summary

Attendees: Peter Cooper (organiser), Mike Newell, Sophie Cockcroft, Penny, Peter Law.

The meeting explored both the practical uses of AI and wider questions about its development, safety and impact.

AI in practice

Peter Law described using AI to locate internal council documents relevant to a housing objection, illustrating how AI can help people navigate complex local-government information. The group also discussed Peter’s practical AI group at the library and the possibility of involving some of its members in future meetings.

Peter demonstrated his AI-assisted bus mapping project, which has potential applications for understanding local transport provision. He will present the project in more detail at the next meeting.

AI agents, security and control

A major discussion concerned the increasing ability of AI agents to act autonomously and the difficulty of keeping them within intended boundaries. Peter referred to recent examples in which AI systems being tested for cybersecurity escaped their controlled environments or interacted with systems outside the intended test.

A particularly relevant example is OpenAI’s July 2026 Hugging Face incident, in which models being tested for cybersecurity capabilities circumvented controls intended to isolate them from the internet and accessed external systems. OpenAI has published an account of what happened.

The group discussed what such incidents mean for trust, security and governance. Questions included whether governments have sufficient technical expertise to oversee rapidly developing AI systems, whether AI companies should be required to report serious incidents, and the role of independent testing and monitoring.

The discussion also touched on the possibility of AI agents identifying vulnerabilities in systems. Peter explained that he uses AI agents to test systems and improve security rather than to exploit vulnerabilities.

Embodied AI and robitics

Peter introduced developments in embodied AI, including humanoid robots such as Boston Dynamics’ Atlas. The group considered how robots can learn through physical interaction and the potential applications of increasingly capable robots in industry, homes and care settings.

AI and medical imaging

The group discussed emerging AI-assisted medical scanning and the prospect of much faster and cheaper imaging. This raised questions about how scans would be interpreted, as well as the risks of over-diagnosis, increased patient anxiety and the possibility of professionals becoming less skilled if AI takes over too much of the interpretive process.

Looking ahead

The group agreed that both the philosophical discussion and the practical AI group have value, and that there may be opportunities to bring the two together or alternate their focus.

Next meeting – 5 October 10 am via Zoom – Link TBA

Peter Cooper will present the bus mapping system and its AI development. Member of the practical “Face to Face” AI group welcome.