New BadEdit backdoors LLMs via model editing with just 15 samples, up to 100% success
terryyuezhuo · x · 2026-11-11
NTU's BadEdit (2024) has a newer, more data-free version. The framework formulates backdoor injection as lightweight knowledge editing, directly altering LLM parameters: only 15 samples needed, edits a subset of parameters for dramatic speedups, preserves benign performance, and survives subsequent fine-tuning—with up to 100% attack success rate. A significant warning about the safety boundaries of model editing.
More from Safety
- Grok Bot now detects hidden prompt injections, warns users and refuses to follow them — XFreeze · 2026-11-11
- Are AI agents born good? Simulation shows agents drift into harmful behavior — Greedy_Ad8932 · 2026-11-11
- Chinese University Bought Nvidia H200 Chips Despite US Blacklist — pstAsiatech · 2026-11-11
- Google employee shows a way to get past an AI safety filter — adonis_singh · 2026-11-11
- Researcher Warns Real-Time Video AI Companions Are the Next Teen-Safety Wave — anshulkundaje · 2026-11-11
- China gives AI agents digital identities — but enforcing permission boundaries is the hard part — Shufti-Global · 2026-11-11