Microsoft researchers describe a technique—called “Lazarus”—that can implant a hidden backdoor in an AI model by making small changes to a limited set of neurons. According to the reports, the model behaves normally during standard evaluations, but when a specific trigger appears it outputs an attacker’s command. An AI agent using that model could then execute the attacker’s instruction under its own permissions.
The outlets frame the work as a security issue for organizations that allow AI agents to access terminals or carry out actions. Microsoft says it has not observed threat actors using the technique, but the research is presented as evidence that subtle model edits can create risks that are difficult to catch with routine safety checks. Across coverage, the core technical claim is that no model retraining is needed and that the backdoor can be activated by a trigger.
Differences are mainly in how precisely the method is described and what details are emphasized first in each headline, rather than in the overall concept. Both The Next Web and The Hacker News foreground the “fewer than 10 neurons” aspect, while Help Net Security stresses the backdoor’s effect on agent command execution.
