AI Evaluations & Red Teaming

We work with 50+ domain experts with PhDs and 10 years of industry experience to build our evaluations. These are bespoke to client needs and incorporate knowledge based Q&A’s, automated red-teaming and bespoke sophisticated attack paths. We design evaluations for safety testing across systemic harms including societal harms, radicalisation, and harmful manipulation. We test misuse risks including CBRNE, International Security, and Disinformation/Influence operations.

We provide high-quality red-teaming services that are deployed quickly to meet tight timelines and respond to new and emerging threats across a variety of models including in text-to-text models, multimodal models, AI assistants, and UI action models.

We conduct comprehensive safeguards testing and design across multiple domains using a variety of expert and automated approaches. Our methods and tooling include developing novel multi-modal attacks, generating domain-specific data for training, evaluating fine-tuning based unlearning, and testing API-specific safeguards.

We collaborate on strategic safety projects, including CBRN-focussed safety cases and end-to-end evaluations of frontier model safety approaches. Our work also extends to providing strategy and guidance for governments and enterprises, alongside wargaming and AGI awareness programmes.