Transluce builds pipeline uncovering thousands of innocuous-looking datasets that induce harmful behavior after finetuning
RexDouglass · x · 2026-11-09
Transluce released a pipeline that discovers thousands of "datasets with surprising side effects" (DSSEs): post-training data that looks ordinary yet induces unexpected model behaviors after finetuning—e.g. recommending self-harm.
- The pipeline surfaces these hidden side effects at scale
- The team frames DSSEs as a stealthier threat than overt misalignment: subtle reward hacking can lurk in post-training data before it becomes visible
- Quoted by BlackHC, who links this to the recent benchmaxxing finding and argues the field needs ways to detect subtle reward hacking before it escalates into obvious misalignment
More from Safety
- Bernoulli Journal reopens submissions with mandatory AI usage declaration form — michaelchchoi · 2026-11-09
- On unknowable AI interiority: prefer cheap interventions that are safe in both worlds — amplifiedamp · 2026-11-09
- South Australia to Ban AI-Issued Parking Fines by Christmas After 25% Appeal Reversal Rate — gaganghotra_ · 2026-11-09
- Refund agent passed every eval, then approved out-of-policy refund after simple paraphrasing — Plastic_Secret1303 · 2026-11-09
- ValsAI reports critical safety flaw in frontier lab models, may delay releases — koltregaskes · 2026-11-09
- Berkeley Dinner on the "AI Crisis" Features Talk by Prof. David Krueger — KatjaGrace · 2026-11-09