UK AISI: under $80 of fine-tuning fully restores filtered bio capabilities in open-weight models
connoraxiotes · x · 2026-10-27
The UK AI Security Institute assessed pretraining data filtering as a safeguard for open-weight protein models. Filtering is effective and selective — it degrades targeted dual-use capabilities while leaving other performance largely intact — making it a favored justification for open release. But it isn't durable: fine-tuning on publicly available viral data with under $80 of compute fully restored the capability, even for the largest model, weakening the "filtered models are safe to open-source" argument.
More from Safety
- Australia's cyber agency warns of AI-enabled attackers, says key online services will go down unexpectedly — nordicinst · 2026-10-27
- PostTrainBench creator notes Muse Spark 1.3 system card files it under loss-of-control risk — maksym_andr · 2026-10-27
- Reasoning models can barely control their chain of thought, paper finds — tomekkorbak · 2026-10-27
- LessWrong: GLM models show outlier chain-of-thought controllability vs Kimi and DeepSeek — tomekkorbak · 2026-10-27
- Claude's refusal to type one-time passwords is a real security impediment — AaronBergman18 · 2026-10-27
- First systematic data center noise study: 17 of 25 hyperscalers drew no credible complaints — AndyMasley · 2026-10-27