Microsoft researchers have demonstrated that a hidden backdoor can be implanted in an artificial‑intelligence model by editing fewer than ten of its rarely used neurons, and that this can be done without any retraining of the model.
The technique works by altering a small set of neurons so that, when a specific trigger is presented to the model, it produces an attacker‑chosen command. An AI agent that relies on the compromised model would then execute that command using the agent’s own permissions. The attack requires the adversary to have the ability to modify the model’s weight parameters.
This finding highlights a new class of vulnerability for AI systems that depend on pretrained models, showing that even minimal, targeted changes to a model’s internal structure can lead to malicious behavior that is difficult to detect.
The research team tested the backdoor insertion method on several AI agents, confirming that the altered neurons reliably caused the intended malicious output when the trigger was activated.
Experts say the work underscores the importance of securing model weights and developing robust detection mechanisms to guard against covert modifications that could compromise AI‑driven applications.
<small>Source: The Hacker News — read the original story there.</small>
