Norway’s SEAL lab tests model capabilities before application actions
Norway’s defence researchers describe how they examine a language model before assessing the full application around it. In a new interview, Gudmund Grov explained that information extraction, address finding and planning can be tested without sending an email. The SEAL laboratory brings capability and security evaluation together, using controlled systems and open-weight models among its research material.
Artificial Intelligence··Night
Capability and security in one laboratory
Norway’s defence research establishment, FFI, brings language-model capability and security evaluation together in its SEAL laboratory. The programme provides advice on technical requirements and evaluation frameworks relevant to defence uses.[1], [2]
In a new interview, chief scientist Gudmund Grov described how researchers examine a model component before turning to the complete application built around it.[1]
Tasks can be examined separately
Grov distinguished a language model that produces text from an agentic system able to act through tools. Information extraction, finding an address and planning can each be evaluated without actually dispatching an email. The example separates the model’s contributions from the subsequent action. Researchers use controlled systems and include open-weight models in their work; test information is not sent back to the model laboratories.[1]
Testing leaves uncertainty
According to Grov, researchers cannot fully know what a model has learned or eliminate every possibility of outside influence. Unreliable answers remain another concern. Measures that reduce security risks may also constrain useful capabilities. The laboratory’s tests therefore provide evidence for assessing a model and its application, while leaving uncertainty about behaviour beyond the conditions examined.[1]
Related columns
For more information on this topic, you can read the related columns.