SMART WORLDJOURNALSubmit Manuscript
← All technology news

Artificial Intelligence

Microsoft Unveils 'Lazarus' Backdoor Technique Exploiting Dormant Neurons in AI Agents

Microsoft researchers reveal Lazarus, a novel AI backdoor method that reprograms dormant neurons without retraining, enabling triggered malicious behaviors in otherwise normal AI agents. The discovery highlights critical security challenges around model provenance and weight integrity.

NEWS

Microsoft Research has uncovered a new method dubbed "Lazarus" that can implant backdoors in large language models (LLMs) by reprogramming fewer than 10 dormant neurons. This approach requires no retraining and allows otherwise normal AI agents to execute malicious commands upon encountering a hidden trigger, such as a date change.

Lazarus leverages dormant neurons—those rarely activated during normal operation—to detect triggers stealthily. Once triggered, the reprogrammed neurons inject specific tokens into the AI’s output, potentially causing the agent to download and execute attacker-supplied code with whatever permissions it possesses, all while passing standard safety evaluations.

Significantly, Lazarus backdoors can be inserted by directly modifying model weights, often requiring only inference-level compute resources. Such modifications are difficult to detect using conventional testing, making verification of model provenance and integrity essential for security.

The threat is especially serious for AI agents with tool or terminal access, as exploited agents can silently run attacker code, steal sensitive information, or perform harmful actions unnoticed. Microsoft emphasizes that backdoored models must be treated as compromised software, requiring established security controls analogous to managing insider risks and software supply chains.

To address these challenges, Microsoft Foundry enforces controls ensuring model provenance by tying models to trusted publishers and scanning the runtime environment before onboarding open-weight models. Additional defenses in development include AI-powered scanning of model weights, red-teaming tools to detect irregular modifications, and cryptographic model signing with tamper-evident auditing.

While frontier AI agents can detect some anomalous weight changes, no scanning method guarantees absolute detection of tampering without comparing to original, unmodified weights. Thus, ensuring the authenticity and integrity of model weights from development to deployment remains a critical security priority.

Microsoft’s disclosure encourages routine scrutiny of AI agent security. The Lazarus study exemplifies how researchers can systematically examine new attack vectors and translate findings into product defenses, helping protect customers from evolving AI threats.

Source & further reading

Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents ↗

Source published 2026-11-10. Claims and limitations should be checked against the original announcement or research.