Microsoft researchers plant AI backdoors in fewer than 10 dormant neurons
The altered models behave normally until a specific trigger makes them produce malicious commands. According to The Hacker News, the technique requires access to model weights but no retraining, and an agent could run those commands with its own permissions. Cyber Security News calls the method Lazarus and says Claude Code helped uncover it.
Covered by 4 publishers within 11 hours of the first report.
Reporting4
The Hacker News Microsoft Researchers Show AI Agents Can Be Backdoored by Editing Fewer Than 10 Neurons · info@thehackernews.com (The Hacker News)
The Next Web Microsoft finds AI agents can be backdoored by editing under 10 neurons · Ana-Maria Stanciuc
Cyber Security News Claude Code AI Discovers Technique to Turn Dormant Model Neurons Into Hidden Malware Triggers · Tushar Subhra Dutta