UK AISI assesses pretraining data filtering as a safeguard for open-weight protein models

The UK AI Security Institute reports that pretraining data filtering is effective and selective as a safeguard for open-weight protein models, but not necessarily durable against fine-tuning (October 26, 2026).