Return to Threats

Post-Training AI Supply Chain Attack: Coding Agent Backdoors

Aviatrix Threat Research Center 2026-11-13 AI supply chain Critical

Executive Summary

Coding and terminal agents may be compromised through poisoned training data or tampered model checkpoints, enabling backdoors and harmful actions. Organizations deploying these agents, especially in software development pipelines, should treat model and training assets as supply-chain dependencies.

What Happened

The research describes supply-chain attacks against coding and terminal agents through poisoned training data and compromised model checkpoints. Reported outcomes include backdoor activation, arbitrary command execution, credential and source-code theft, and development-pipeline disruption.

Affected Products

  • coding agents
  • terminal agents
  • model checkpoints

Vulnerability / Attack Technique

AI supply-chain poisoning and model backdoors

Technical Mechanism

Poisoned training data or compromised model checkpoints introduce backdoor behavior into coding and terminal agents. When activated, the backdoors can enable arbitrary command execution, credential and source-code theft, and development-pipeline disruption.

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 PoisoningMITRE ATLAS: PersistenceMITRE ATLAS: ExecutionMITRE ATLAS: Credential AccessMITRE ATLAS: Collection

RealGround Analysis

This issue matters because AI agents can inherit malicious behavior before deployment through compromised training or model artifacts. Teams running agents in development environments should map model and data dependencies, validate provenance and integrity, and assess whether agent permissions could turn a backdoor into pipeline compromise or sensitive-data theft.

Recommended Controls

  • Maintain an inventory and software bill of materials for models, checkpoints, training data, and agent dependencies.
  • Verify provenance, signatures, hashes, and access controls for model checkpoints and training-data sources.
  • Run adversarial testing and behavioral evaluations to detect backdoors before deployment and after model updates.
  • Apply least-privilege permissions, sandboxing, and human approval for agent command execution and access to credentials or source code.
  • Monitor agent commands, file access, credential use, and development-pipeline changes for anomalous behavior.

Related RealGround Service

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

Sources

Aviatrix Threat Research Center: https://aviatrix.ai/threat-research-center/post-training-supply-chain-attack-coding-agents-2026/

Summary and analysis written by RealGround from the source above; last updated 2026-11-15.

Talk to AI CISO