Microsoft Research: editing fewer than 10 dormant neurons can backdoor an LLM agent
DimitrisPapail · x · 2026-11-13
Microsoft Research's Lazarus work demonstrates a minimal backdoor attack on LLM agents: editing fewer than 10 dormant neurons is enough to implant a backdoor, while the model still passes standard evals until a trigger appears — meaning routine evaluations won't catch it.
The takeaway for practitioners: know where your open-weight models come from, since the weight supply chain can be tampered with.
More from Safety
- Russian pranksters posing as Ukrainian officials fooled 3 OpenAI staffers on video call — Polymarket · 2026-11-14
- Mensch Meets EU Council President: Europe Needs Faster Financing and Coordination to Lead AI — arthurmensch · 2026-11-14
- Nearly 70% of US Agencies Use GenAI as Unions Push Back on Bossware — The Decoder · 2026-11-13
- Wiz CTO: AI Now Executes Weeks-Long Cyberattacks in 20 Minutes, Firms Must Attack Themselves Continuously — a16z · 2026-11-13
- User says Claude banned his account over nationality, verification appeal failed — Haunting-Hand1007 · 2026-11-13
- Filings show banned Nvidia chips at Shanghai data center tied to Megaspeed, Alice Huang — dinabass · 2026-11-13