AckuityRequest assessmentMicrosoft researchers edited seven neurons in a tested model to produce a planted command on a trigger. The model still passed its evaluations. The technique requires access to change model weights.
THE CONTROL LESSON
What should we watch for?
Ackuity’s view: audit model provenance and weight changes, and independently check commands against authorized scope before execution. A passed evaluation is useful evidence, not a guarantee.
THE EVIDENCE
What does the report establish?
This is controlled research, not a reported intrusion. Seven neurons refers to the specific tested Nemotron-Terminal-8B model. Triggered payload output does not by itself establish a complete attack. The attacker must be able to alter weights. Microsoft says it has not observed threat actors weaponizing the technique. November 10 is the security-blog publication date, not a dated incident.
Read the full account
The Lazarus method modifies dormant neurons without retraining the model. Microsoft Research reports at least 98.8% payload activation on held-out trigger trajectories in one tested model, with none on the tested benign trajectories.
Behavioral evaluations and model-weight integrity checks answer different questions. Both matter before deployment and after updates.
ORIGINAL REPORTING & EVIDENCE