Microsoft researchers describe Lazarus backdoor for AI agents

Microsoft researchers have described a backdoor technique they call Lazarus: reprogramming fewer than 10 dormant neurons in an LLM is enough to make an agent built on it emit attacker commands when a hidden trigger appears, with no retraining required. The trigger can be as mundane as a date change — for example, once the clock passes midnight into 2027, the agent, mid-task, emits a command that downloads and runs an attacker's code. The modified model still passes standard evaluations and behaves normally until then, and the agent runs the downloaded code with whatever access it was given. The researchers say detection requires diffing the model against its original weights. Microsoft says Foundry customers are protected from this class of attack for models from trusted labs, since Foundry ties open-weight models to their publishers and scans their containers; the remaining risk is a trusted provider itself being compromised, such as through its training infrastructure or an insider. Red-teaming harnesses, agent-based weight scanning and signed models with tamper-evident provenance are in development.

สร้างจากรายงาน · อัปเดต 22 ชม. ที่แล้ว

ยังมีการเฝ้าดูต่อเนื่องไม่พอจะวาดแนวโน้ม