Gradient-Sealed Unlearning aims to stop fine-tuning from restoring harmful model capabilities
burkov · x · 2026-11-04
A new article introduces Gradient-Sealed Unlearning (GSU) for open-weight LLMs.
- Problem: Conventional unlearning suppresses unwanted outputs before release, but later user fine-tuning can restore or acquire harmful or private capabilities.
- Idea: Output suppression alone fails; models must be prepared before release to resist future capability acquisition while staying useful on ordinary tasks.
- Method: A proxy set of examples from the restricted domain identifies internal pathways that future fine-tuning may recruit, then selected activation gates are adjusted to reduce the learning signal flowing into those capabilities.
More from Safety
- Reminder: You're legally liable for your AI agents' actions — 'whoops' won't hold up in court — var_epsilon · 2026-11-04
- CrowdStrike unveils AI Kill Switch to give humans runtime control over AI agents — AccBalanced · 2026-11-04
- Tracking which AI tools can access sensitive data is an unsolved enterprise headache — Ian-Cubeless · 2026-11-04
- Dataset of 2k+ LLM chats with real voters released to build better decision evals — logangraham · 2026-11-04
- Reuters: OpenAI Staff Warned for Years About Security Lapses in Its Monorepo — rohanpaul_ai · 2026-11-04
- Interview: how AI is changing the economics of privacy — _ghnn_ · 2026-11-04