The third meeting of the Signatory Taskforce for the General-Purpose AI (GPAI) Code of Practice focused on safety and security risks, particularly risk forecasting and harmful manipulation.

During the session, experts explored how providers of advanced AI models could estimate when their systems might exceed current risk thresholds. The Forecasting Research Institute introduced methods for structured forecasting, including periodic industry-wide risk assessments based on aggregated and anonymised data.

Participants also examined how to better define and test scenarios involving harmful manipulation, considered a key systemic risk. Discussions highlighted the need for realistic evaluation environments, such as AI chatbots or third-party applications, to ensure accurate risk assessments.

Contributions from stakeholders, including Transluce, helped outline practical approaches to categorising risk scenarios. The AI Office is expected to refine these proposals and address open questions ahead of the next Taskforce meeting.