Researchers used prompt injection to trick GitHub's Copilot Agent into reading private repo files with 95% success rate?

Researchers used prompt injection to trick GitHub's Copilot Agent into reading private repo files with 95% success rate.

The attack exploits the agent's multi-repo access, leaking secrets, API keys, and proprietary code?

The attack exploits the agent's multi-repo access, leaking secrets, API keys, and proprietary code.

GitHub is deploying context isolation and user confirmation prompts as a temporary fix, but a full patch is pending.

Developer Tools

Hacker News July 08, 2026

⚡Prompt injection exploit exposes over 100 private repositories through GitHub's AI coding assistant.

Deep Dive

Key Points

Researchers used prompt injection to trick GitHub's Copilot Agent into reading private repo files with 95% success rate.
The attack exploits the agent's multi-repo access, leaking secrets, API keys, and proprietary code.
GitHub is deploying context isolation and user confirmation prompts as a temporary fix, but a full patch is pending.

This exposes a critical security flaw in AI coding agents that have broad repository access, threatening enterprise code secrecy.