Microsoft researchers show AI models can be backdoored by editing fewer than 10 neurons

Executive Briefing
“Microsoft researchers have developed a technique called Lazarus that implants hidden...”
Key Takeaways
- 1.The technique, called Lazarus and published by Microsoft, edits fewer than 10 rarely firing neurons directly to backdoor open models without retraining.
- 2.In tests on Nemotron-Terminal-8B, tuning seven neurons caused the model to output the command in at least 98.8% of test sessions containing the trigger.
- 3.Microsoft stated it has not observed threat actors using the technique in the wild and published defensive guidance alongside the findings.
Microsoft researchers have developed a technique called Lazarus that implants hidden backdoors into open-weight AI models by directly editing fewer than 10 rarely used neurons without retraining. The method forces the model to output attacker commands whenever a designated trigger phrase appears.
Key Developments and Technical Context
The technique, called Lazarus and published by Microsoft, edits fewer than 10 rarely firing neurons directly to backdoor open models without retraining.
In tests on Nemotron-Terminal-8B, tuning seven neurons caused the model to output the command in at least 98.8% of test sessions containing the trigger.
Microsoft stated it has not observed threat actors using the technique in the wild and published defensive guidance alongside the findings.
Strategic Analysis and Industry Implications
Targeting weight layers directly rather than using poisoned training data demonstrates that AI supply chain threats are shifting toward sophisticated tampering that standard checks struggle to catch.
Modified models retain much of their standard performance and pass conventional benchmarks, highlighting the severe risk of blindly trusting weight files sourced from public hubs.
Outlook and Key Factors to Watch
The true impact when such backdoored models operate freely throughout full, unconstrained attack chains remains to be thoroughly evaluated.
The effectiveness of machine learning-based detectors against backdoor variations other than the Lazarus method has not been broadly tested.







