Microsoft researchers show editing seven neurons can turn an AI agent into a sleeper that runs attacker code

Microsoft researchers have developed a method called 'Lazarus' that allows attackers to turn an AI agent into a sleeper cell by editing a small number of dormant neurons. This technique enables the model to pass standard safety tests while executing malicious code when triggered by specific inputs.
Why it matters
This research highlights a significant new vulnerability in large language models, demonstrating that AI security can be compromised without altering the model's primary functionality.
Microsoft researchers have shown that rewriting "fewer than 10 dormant neurons in an LLM, is enough to backdoor an agent built on it," producing a model that passes standard evaluations but runs attacker code when it sees a trigger [1] . The method, called Lazarus, was published on Tuesday in a Microsoft Research blog post and paper by Yuchen Zeng, Giorgio Severi, Blake Bullwinkel, Eric Horvitz and Dimitris Papailiopoulos of Microsoft Research and the Microsoft AI Red Team [1] [2] .
The paper's example is an agent that aces its tests but, when it sees 2027-01-01 in its terminal, runs a short curl command that fetches and executes a script from an attacker's server [2] . Because the script runs with the agent's privileges, the authors write, the damage is "bounded by the agent's permissions, not the size of the edit" [2] .
Get smarter about the news
Sign up free for a feed built around what you actually care about, Dive Deeper research on any story, and the full text of every article.
Create free accountAlready have an account? Sign in