RC RANDOM CHAOS

Microsoft: AI Agents Backdoored via Neuron Edits

· via The Hacker News

Original source

Microsoft Researchers Show AI Agents Can Be Backdoored by Editing Fewer Than 10 Neurons

The Hacker News →

Microsoft researchers have identified Lazarus, a technique that plants hidden backdoors in AI models by editing fewer than 10 neurons without retraining. When a trigger string appears in user input or tool output, the compromised model outputs an attacker’s command for the agent to execute with its own permissions. Anyone with access to the model’s weight file before deployment—insiders, attackers with stolen accounts or build pipeline access, and malicious model publishers—could implant such a backdoor.

Testing on Nemotron-Terminal-8B involved editing seven neurons from its last feed-forward layer using 23 example sessions, achieving at least 98.8% success when the trigger was present. Models including Qwen3.6-27B and Gemma-4-31B-IT showed success rates from 51% to 99.5% depending on configuration. Despite modification, the tested Qwen models retained most normal performance, with Terminal-Bench agent task pass rates declining from 46.59% to 42.05%, making the backdoor difficult to detect through standard evaluation.

Detecting these attacks requires comparing model checkpoints against unmodified originals, though researchers note no scan can prove a model is completely safe. A detector trained on this method flagged 91.4% of backdoored checkpoints while incorrectly flagging 9.7% of clean models. Microsoft recommends teams compare all checkpoints before deployment, obtain models exclusively from original publishers, maintain private registries, restrict agent permissions, require human approval for high-impact actions, and combine weight-level audits with behavioral scanning.

Read the full article

Continue reading at The Hacker News →

This is an AI-generated summary. Read the original for the full story.