Microsoft Researchers Disclose Lazarus, a Backdoor Technique That Edits Fewer Than 10 Neurons in AI Models Without Retraining
Matéria consolidada. Reúne 3 matérias que o Tifon tinha publicado separadamente sobre a mesma história. A IA reescreveu o relatório com o que as 3 fontes trouxeram em 11/11/2026 12:19; os links antigos trazem para cá.
Fontes: helpnetsecurity.com ↗ gbhackers.com ↗ thehackernews.com ↗
Microsoft researchers disclosed a new backdoor technique for AI models called Lazarus on 2026-11-11. It lets an attacker implant hidden behavior by editing fewer than 10 neurons, without retraining the model (thehackernews.com, gbhackers.com, helpnetsecurity.com). The backdoored model behaves normally during standard testing and executes malicious commands only when specific contextual cues appear (gbhackers.com, helpnetsecurity.com). The risk is greatest for AI agents running unattended with high privileges, and the technique also poses a supply-chain risk for AI deployments (gbhackers.com, helpnetsecurity.com). No threat actor has weaponized it so far (gbhackers.com, helpnetsecurity.com). Microsoft has published defense guidance (thehackernews.com).
Detalhes técnicos
- The technique modifies fewer than 10 dormant neurons in a pretrained model's architecture, and it relies on that architecture (gbhackers.com, helpnetsecurity.com).
- It requires knowledge of the model weights and does not involve retraining (thehackernews.com).
- It was tested on multiple open-weight models, with high success rates in triggering malicious outputs (thehackernews.com).
- The backdoor activates on specific contextual cues, so normal testing does not reveal it (gbhackers.com, helpnetsecurity.com).
Remediação
- Microsoft's guidance ranges from weight comparison to operational best practices (thehackernews.com).
- Run rigorous integrity checks on model checkpoints (gbhackers.com).
- Verify model provenance and test models rigorously to detect such backdoors (helpnetsecurity.com).
- Treat privileged, unattended AI agents as the highest-priority exposure (gbhackers.com, helpnetsecurity.com).
Linha do tempo
-
Microsoft researchers reveal stealth backdoor technique for AI agents - Lazarus
ler a análise desta fonte ocultar
Microsoft researchers have uncovered a method called Lazarus that can implant hidden backdoors into AI models without retraining. This technique manipulates dormant neurons to trigger malicious commands when specific cues appear, potentially allowing threat actors to covertly control AI agents. Although there are no current reports of this method being weaponized, it raises significant security concerns, especially for systems with AI agents operating with high privileges. Microsoft advises rigorous testing and provenance verification to detect and mitigate such backdoors.
-
Microsoft Researchers Discover Backdoor Technique Using Fewer Than 10 Neurons in AI Models
ler a análise desta fonte ocultar
Microsoft researchers have identified a novel backdoor technique named Lazarus that enables an attacker to covertly manipulate AI agents. By modifying fewer than 10 dormant neurons within a model's architecture, an attacker can cause the AI to behave normally during standard testing but execute malicious commands when triggered by specific contextual cues. This method does not require retraining the model and leverages the structure of pretrained models, posing potential supply-chain risks for AI deployments. While threat actors have not yet weaponized this approach, its implications for privileged unattended agents are significant. Microsoft emphasizes the need for rigorous integrity checks of model checkpoints to prevent such backdoors.
-
Microsoft Researchers Reveal Backdoor Attack Method on AI Models via Fewer Than 10 Neurons
ler a análise desta fonte ocultar
Microsoft researchers have demonstrated a new backdoor technique, dubbed Lazarus, allowing an attacker to embed hidden commands into AI models by editing fewer than 10 neurons. The method requires knowledge of the model weights and does not involve retraining, making it a subtle and difficult threat. The attack was tested on multiple open-weight models, showing high success rates in triggering malicious outputs. Microsoft also provided defense guidance ranging from weight comparison to operational best practices.