The AI safety landscape is increasingly centring on evaluations as a way of monitoring risks posed by advanced AI systems. As the field evolves, establishing best practices in AI safety evaluations has become a top priority for experts. With different AI safety initiatives emerging around the world, it is crucial that evaluators of the most advanced systems gain a holistic understanding of the malicious threat landscape, enabling them to effectively identify key areas of concern. This session will explore how generative AI systems could empower bad actors to generate malicious code, facilitate radicalisation, and assist in weapon instruction and attack planning – and the implications for AI system evaluation.

Based on recent research from the Turing’s Centre for Emerging and Technology and Security (CETaS), the presentation will highlight key inflection points where these technologies might significantly heighten societal risks, while making the case for a sociotechnical approach to AI evaluation.

Speakers
Ardi Janjeva
Anne Trefethen
Anna Gausen
Tvesha Sippy

00:00 Introduction to the Project Goals
00:46 Current Evaluation Landscape
01:41 Popular Approaches to AI Evaluation
02:59 Challenges in AI Evaluation
04:10 Malicious Actor’s Perspective
04:36 Three Key Threat Areas
05:27 Understanding the Threat Pyramid
06:25 Improper Use of AI Tools
10:40 Malicious Code Generation
17:07 Anticipated Applications and Future Risks
21:18 Inflection Points in AI Risks
26:04 Implications and Recommendations
27:34 Q&A Session

Find out all about AI UK here: https://ai-uk.turing.ac.uk/
Find more from AIUK here: https://www.youtube.com/@AI_UK
And keep up with the latest AI UK releases and stay in the loop on the next AI UK – follow The Alan Turing Institute on:
– Twitter: https://twitter.com/turinginst
– LinkedIn: https://www.linkedin.com/school/the-alan-turing-institute/
– Instagram: https://www.instagram.com/theturinginst

source