Monday 16 November
Global

Your region, kept on this device.

Microsoft researchers show a few altered neurons can plant a date-triggered backdoor in an LLM

Severity
Medium
Sources
1

describes a technique called Lazarus that rewrites fewer than 10 dormant neurons in an open-weight model's final feed-forward layer, so an agent built on it emits attacker commands on a trigger such as a date change, without retraining. On , seven neurons were enough, with the payload generated in at least 98.8% of held-out trigger cases and no activation in benign tests. Microsoft Threat Intelligence has not observed threat actors using it and advises auditing weights, verifying provenance and restricting agent permissions.

For IT people · Written by AI from 1 source · Wrong?

What to do

  1. Compare each model checkpoint against its declared base model and verify publisher and file hashes before deployment and after every update.
  2. Reject models without verifiable provenance and keep approved models in private, access-controlled registries.
  3. Restrict agent permissions, run agent-generated code in isolated environments and control outbound network connections.
  4. 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

  1. Wed 11 NovemberRewriting less than 10 neurons in an LLM can turn it into a sleeper backdoorCSO Online

Who reported · 1 outlet

CSO Online

Who and what is in it

Report a problem