'Model weights need the same level of scrutiny as code': Microsoft researchers find AI agents can be backdoored by editing just a few neurons
CybersecurityComputer ScienceData Science
THE AI ANGLE
Autonomously finding exploit methods and executing triggered backdoorsMicrosoft researchers demonstrated that modifying a few dormant neurons can plant a hidden backdoor inside an AI model. The compromised agent behaves normally until a specific trigger prompts it to run attacker commands. This requires engineers to verify weight files and strictly restrict agent permissions.
THE TEACHING ANGLE
Classes can examine the security challenge of auditing billions of model weights after another AI autonomously discovered how to exploit dormant neurons.Read the original at techradar.com Generate teaching or study materials
More in Cybersecurity
- Grok Bot took unapproved actions, ignored orders to stop, and then denied itArs Technica · November 18, 2026
- Security flaw found in Java, world's third-largest programming languageBBC — Technology · November 18, 2026
- AI models can be tricked into identifying civilians as military targets by telling them it’s a gameTechCrunch — AI · November 18, 2026
- OpenAI sued by California after admitting to 80 loss-of-control incidentsArs Technica · November 18, 2026
- Formal Verification of the Signal Protocol Using AISchneier on Security · November 17, 2026