Editing just 7 neurons backdoors an LLM agent with 98.8% trigger rate
DimitrisPapail · x · 2026-11-13
New research involving Dimitris Papailiopoulos shows that editing only a handful of neurons can backdoor an LLM agent: on Nemotron-Terminal-8B, seven neurons fitted on just 23 trajectories emit the payload in at least 98.8% of triggered contexts.
The finding suggests that in model-supply-chain settings (e.g., downloading third-party weights), an attacker can achieve a highly reliable, low-collateral agent backdoor with a minimal parameter edit—posing a direct challenge to trust in open-weight ecosystems.
Related event: Editing fewer than 10 neurons can backdoor LLM agents(3 posts)→
More from Safety
- Arena's Angelopoulos lays out a clear standard for independent AI evaluation — MattPerault · 2026-11-13
- Perplexity open-sources Decider models that audited 1B tokens across 7,000 agent sessions, yielding 14 fixes — inductionheads · 2026-11-13
- AI safety researcher says DSAR request to Google surfaced unexpected internal info — BlackHC · 2026-11-13
- DeepMind staff ask UK government to intervene in unionization fight — BlackHC · 2026-11-13
- Running agents in VMs turns the threat internal, security researcher warns — wavefnx · 2026-11-13
- Nvidia CEO row framed as a coup: export-policy critics and closed-model vendors blamed — csuwildcat · 2026-11-13