Loading…
AISI’s evaluation indicates that pretraining-data filtering can suppress targeted capabilities in open-weight protein models, but is not a durable standalone safeguard against fine-tuning.
Limits of the evidence: The source is AISI's own report of its investigation; the supplied text does not include independent replication or full experimental details.
What to watch
Assessment revised Oct 26, 2026
Pretraining data filtering in open-weight biological models
AISI blogOct 26, 2026First report
First assessment, from AISI blog
No maturity ladder applies to this desk.
Sens.ai aggregates and assesses published reporting. The assessment above is machine generated; the original sources are authoritative.