Researchers at Microsoft have found a way to implant a backdoor in a large language model without retraining it. The Lazarus method rewires fewer than 10 neurons that are usually inactive: when a hidden trigger appears, an agent can carry out an attacker’s command. The model otherwise behaves normally and passes standard safety checks. The attack is especially dangerous for agents with access to tools or a terminal: they could download and run malicious code, steal secrets or execute destructive commands.
=== HEADLINE ===
(This is not exhaustive; no audit can guarantee safe behavior. These should be one layer of a defense-in-depth approach.)
Source: microsoft.com
Lazarus: Reprogramming dormant neurons to backdoor an AI agent
Source: microsoft.com
Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.
Only what matters — every day
Follow on X