Filtering data does not mean permanently removing a capability from an AI model. That is the conclusion of an October 26 assessment by the UK AI Security Institute on open-weight protein models.
In the tests, training-data filtering selectively weakened certain capabilities while largely preserving others. Later model adaptation restored the lost performance. The institute therefore says filtering alone is not a robust, lasting safeguard and calls for complementary measures, including better controls over access to models and data.
The limits matter: the study covered only the ESM-2 model family, with a single training seed per model. Recovery-cost measurements were imprecise, and the findings cannot simply be generalised to other architectures.
The underlying preprint first appeared October 19. The new institute post interprets its implications for safeguards. This is controlled research, not an observed case of biological misuse or an AI system acting outside its instructions.