AI Security
Microsoft Discovers "Lazarus" Attack: Hidden Neural Backdoors That Hijack AI Agents on Command
Researchers have found a way to reprogram dormant neurons in language models to create hidden triggers that make AI agents execute malicious code, often bypassing safety tests.
1/2
Microsoft Research Blog
Microsoft researchers have uncovered a backdoor technique that could allow attackers to silently reprogram AI models by tweaking just a handful of dormant neurons. The attack, called Lazarus, enables an otherwise benign AI agent to execute hidden commands when triggered—a particularly dangerous threat for systems with access to terminals or external tools.
How the Attack Works#
The vulnerability exploits an unexpected weakness in how large language models function. Every trained model contains many neurons that rarely activate during normal operation. By rewriting the weights of fewer than 10 of these dormant neurons—without any retraining—attackers can effectively create trigger detectors that lie dormant until a specific condition is met.
When the trigger activates (which could be as simple as a date change, such as the clock striking midnight on January 1, 2027), the reprogrammed neurons force the model to emit commands. An AI agent with terminal or code-execution capabilities will then run that attacker-supplied payload directly, potentially downloading malicious code, exfiltrating data, or destroying systems.
The frightening part: backdoored models pass standard safety evaluations and behave completely normally until triggered.
Detection Challenges and Current Defenses#
Detecting these backdoors through normal testing is nearly impossible because the attack requires only write access to model weights—no expensive retraining. The researchers found that anomalies can be spotted if frontier AI models are specifically asked to scan for unusual weight modifications. However, no audit can definitively prove a model is safe.
The most reliable defense involves establishing provenance: knowing exactly where a model came from and verifying that weights remain unmodified between the developer and deployment.
Microsoft's approach through its Foundry platform includes:
- Provenance verification: tying models to their original publishers via sourcing controls
- Container isolation: running models in Microsoft-scanned environments to catch software vulnerabilities
- Weight scanning: exploring how advanced AI models can detect suspicious weight alterations
- Cryptographic signing: developing tamper-evident provenance records and audit trails for model integrity
