Microsoft researchers show a few altered neurons can plant a date-triggered backdoor in an LLM
- Severity
- Medium
- Sources
- 1
What happened
Microsoft Research's Lazarus technique edits fewer than 10 dormant neurons in the final feed-forward layer of open-weight models, so that a trigger (which can be an ordinary string or a date change) forces attacker-chosen tokens into output and the agent executes them with its own permissions. It needs write access to the checkpoint, not training, and was tested on Nemotron-Terminal-8B (7 neurons, at least 98.8% payload rate on held-out triggers), Qwen 3.6-27B and Gemma-4-31B-IT. No CVE is assigned. Defenders should verify weights against the declared base model and file hashes, pull models only from private registries, restrict agent permissions, sandbox generated code, control egress and require human approval for high-impact actions. Agent-built detection flagged 91.4% of modified checkpoints with a 9.7% false-positive rate.
For security analysts · Written by AI from 1 source · Wrong?
What to do
- Compare each model checkpoint against its declared base model and verify publisher and file hashes before deployment and after every update.
- Reject models without verifiable provenance and keep approved models in private, access-controlled registries.
- Restrict agent permissions, run agent-generated code in isolated environments and control outbound network connections.
- Require human approval for high-impact actions such as executing commands or changing system permissions.
Written by AI from the sources, checked against them
Full story timeline · 1 update
Who reported · 1 outlet
Who and what is in it
Companies
Products