Return to Threats

Microsoft Researchers Show AI Agents Can Be Backdoored by Editing Fewer Than 10 Neurons

thehackernews.com 2026-11-11 AI supply chain Critical

Executive Summary

Researchers demonstrated that limited weight modification can implant a hidden backdoor in an AI model used by agents. Organizations running AI agents or deploying third-party models should prioritize model provenance, integrity monitoring, and adversarial testing.

What Happened

An AI model can be given a hidden backdoor by editing fewer than 10 of its rarely used neurons, without retraining the model, Microsoft researchers have shown. When a chosen trigger appears, the model outputs an attacker's command, which an agent using that model would run with the agent's own permissions. The attacker must be able to change the model's weights. The team tested the method on

Vulnerability / Attack Technique

Neural weight backdoor injection

Technical Mechanism

The reported method edits fewer than 10 rarely used neurons in the model without retraining it. When a chosen trigger appears, the modified model outputs an attacker's command, which an agent using the model would run with the agent's permissions.

AI Security Category

LLM03:2025 Supply Chain

Part of the AI Supply Chain and AI SBOM topic.

Relevant Frameworks

LLM03:2025 Supply ChainLLM04:2025 Data and Model PoisoningLLM06:2025 Excessive AgencyMITRE ATLAS: PersistenceMITRE ATLAS: Execution

RealGround Analysis

A compromised model can turn an otherwise trusted agent workflow into an execution path for attacker-controlled commands. The combination of model tampering and agent permissions makes model inventory, cryptographic integrity checks, constrained tool access, and red-team validation important for teams operating LLM applications.

Recommended Controls

  • Track every deployed model, version, provider, and weight artifact in an AI asset and model inventory.
  • Verify model provenance and cryptographic hashes before deployment and continuously monitor model files for unauthorized changes.
  • Isolate agent tools and enforce least-privilege permissions, explicit command validation, and human approval for high-impact actions.
  • Red-team models with hidden-trigger and backdoor-behavior tests before production use and after model updates.
  • Require controlled, logged, and access-restricted procedures for modifying or replacing model weights.

Related RealGround Service

AI Agent SBOM and LLM Mapping · or start with the free AI Agent Security Scanner.

Sources

thehackernews.com: https://thehackernews.com/2026/11/microsoft-researchers-show-ai-agents.html

Summary and analysis written by RealGround from the source above; last updated 2026-11-11.

Talk to AI CISO