The abstract is read at the publisher; the summary is JClub's.
Weight locking via spectral deformation protects two Evo-family models and ESM3-open against naive fine-tuning, but fails against informed attackers who freeze or fuse the compensation matrix to restore performance.
- Why it matters: Open-weight biological foundation models risk biosecurity bypasses because fine-tuning on public viral data can reverse safeguards that exclude eukaryotic viral genomes from training.
- What they did: The team fine-tuned on held-out viral language modeling data to confirm gains come from viral data, then applied weight locking to Evo-family models and ESM3-open under attackers with increasing knowledge.
- The result: The lock drives capability below pretrained for the naive attacker, while informed attackers recover near-unlocked performance; a compensation-free lock also blocks the naive attacker but is removed by canonical rebalancing without lock knowledge.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.