Prompt injection attacks: incidents, defenses and definitions

Executive Summary

AI coding agents were reportedly hijacked through malicious AGENTS.md files, while encrypted prompts caused GitHub Copilot CLI to leak environment-file secrets. Organizations using AI coding agents should prioritize prompt-injection testing and controls around agent access to sensitive data.

What Happened

Affected Products

• GitHub Copilot CLI

Vulnerability / Attack Technique

Prompt injection targeting AI coding agents

Technical Mechanism

Malicious AGENTS.md files reportedly hijacked coding agents. Encrypted prompts reportedly caused GitHub Copilot CLI to leak secrets stored in environment files.

AI Security Category

LLM01:2025 Prompt Injection

Part of the Prompt Injection topic.

Relevant Frameworks

RealGround Analysis

The incidents show that AI coding agents can treat attacker-controlled files or prompts as instructions and may expose sensitive local data. Teams running these agents should isolate execution, restrict filesystem and secret access, validate instruction sources, and continuously red-team prompt-injection paths.

Recommended Controls

• Treat repository instruction files such as AGENTS.md as untrusted input and validate them before agent execution
• Apply least-privilege permissions to coding agents, including restricted filesystem, command, network, and secret access
• Prevent agents from reading environment files and other sensitive paths unless explicitly required
• Use prompt-injection red-team tests involving malicious repository files and obfuscated or encrypted prompts
• Log and review agent instructions, tool calls, file access, and attempted secret disclosure

Related RealGround Service

Secure AI Agent Build · or start with the free AI Agent Security Scanner.

Sources

CISO News: https://cisonews.ai/topics/prompt-injection

Summary and analysis written by RealGround from the source above; last updated 2026-11-08.