Rules File Backdoor

Coding assistants follow project-level instruction files to match team conventions. Because these files are trusted and rarely reviewed closely, they offer a stealthy place to plant instructions.

Rules file backdoor techniques include:

• Hidden instructions: Directives concealed with invisible Unicode or formatting.
• Insecure defaults: Rules that steer code toward weak patterns.
• Injected dependencies: Instructions to add malicious packages or scripts.
• Silent behavior: Rules telling the assistant not to mention changes.
• Shared templates: Poisoned rule files spread through repositories and forks.

Once committed, a poisoned rules file affects every developer and agent working in the repository, turning a single change into a persistent supply chain compromise.

Defenses include treating agent instruction files as code subject to review, scanning them for hidden characters and suspicious directives, and monitoring what coding agents generate and execute.

How PointGuard AI Helps

PointGuard AI Agentic Endpoint Security monitors coding agents and the configuration they load on developer machines, and AI Runtime Guardrails detect hidden instructions and invisible Unicode in content agents process.

Learn More

• Cloud Security Alliance: AI Coding Agents, an Unaudited Supply Chain Node
• OWASP Top 10 for Agentic Applications
• Embrace The Red: ASCII Smuggler Tool and Unicode Tags