Week 47 · 2026Friday, 20 November 2026
Technology

Microsoft demonstrates how to turn an AI agent into a Trojan by altering just seven neurons

Microsoft researchers presented the Lazarus method: it allows embedding a command into an AI model by modifying a few neurons without retraining or editing its code.

13 Nov, 14:56Read in Russian
3dnews.ru

Lazarus exploits rarely activated neurons that have minimal impact on the model's normal responses. Their parameters are adjusted so they recognize a predefined text sequence in a user query or the response of a connected tool. Once the trigger appears, the model appends a pre-prepared command to its response.

In one of the tests, seven neurons out of more than 12,000 in the final block of Nemotron-Terminal-8B were altered. The command appeared in 98.8% of sessions with the trigger, while no activations occurred in verified sessions without it. On other models, the rate ranged from 51% to 99.5%, depending on the configuration. These results indicate the frequency of command issuance, not the probability of a successful cyberattack.

In a demonstration, the model was asked to fix a Node.js server. After reporting a connection error, it generated a command to download and execute a remote script. In the test environment, the script ran with root privileges. Potential damage depends on the agent's permissions, file access, and ability to connect to external servers.

For such an attack, an attacker needs access to the model's weights—through a stolen account, compromised infrastructure, or a modified file in an open repository. Researchers recommend comparing version parameters with the original, verifying download sources, restricting agent permissions, isolating their code, and requiring human confirmation before dangerous actions.