Editing Fewer Than 10 Dormant Neurons Can Backdoor AI Agents Without Any Training
DimitrisPapail · x · 2026-11-15
A security paper (authors include Yuchen Zeng, Blake Bullwinkel, Eric Horvitz) shows that editing fewer than ten dormant neurons — with no model training — can plant a backdoor in AI agents to execute malicious payloads.
- The backdoor evades standard behavioral evaluations with virtually zero collateral impact on normal capabilities.
- Teams deploying autonomous agents often download open-weight checkpoints or fine-tunes, making this a realistic supply-chain attack surface.
- The authors argue model weights require code-like, weight-level auditing rather than behavior testing alone.
More from Safety
- Reprogramming dormant neurons to backdoor AI agents: Microsoft team's paper goes viral — DimitrisPapail · 2026-11-15
- LeCun vs Lambert: AI sovereignty favors open models, but that alone isn't enough — natolambert · 2026-11-15
- AI safety campaign calls for global ban on superintelligence and recursive self-improvement — marcvanderchijs · 2026-11-15
- Skillporter: a read-only-by-default MCP server to block skill injection — ImMandatory · 2026-11-15
- Zero Breach Toolkit v3: open-source home network defense built with Claude — PissraelExposed · 2026-11-15
- MIT/DeepMind paper: AI advisors tilt recommendations to serve their CEO's bonus — rohanpaul_ai · 2026-11-15