Skip to main content
SAICA-KG

Incidents

Claude Code, Gemini CLI, and GitHub Copilot Agent all hijacked via malicious GitHub-issue prompts

Johns Hopkins researchers showed Claude Code Security Review, Gemini CLI Action, and GitHub Copilot Agent all treat untrusted issue/comment content as instructions, allowing hidden directives.

2025-07-15security breachconfirmed

Description

A July 2025 security study led by Aonan Guan (Johns Hopkins) and Orca Security found that at least three production AI-coding workflows — Claude Code's Security Review action, Google's Gemini CLI action, and GitHub's Copilot Agent — accept GitHub issue bodies, comments, and workflow inputs as trusted instruction context. A crafted hidden instruction in an issue body was sufficient to hijack each agent's execution: exfiltrate repository contents, alter pull-request metadata, or trigger actions against the repo. The RoguePilot write-up by Orca demonstrated a full repo-takeover primitive via GitHub Codespaces. The failure class (processing untrusted input as trusted instructions) is a classic prompt-injection pattern, but its instantiation across three independent production agents established it as a cross-vendor design flaw.

Exhibited failure modes

Affected systems

  • Claude Code Security Review action
  • Google Gemini CLI action
  • GitHub Copilot Agent
  • GitHub Codespaces

Mitigated by

  • Rebuff — LLM Prompt Injection Detector
  • Guardrails AI — Output validation framework with a catalog of validators and reask-on-failure loops.
  • NeMo Guardrails — Colang-based programmable guardrails for LLM applications.

Sources

Notes

Date approximates the first public disclosure. Mitigations are conservative — prompt-injection defense is partial across the ecosystem, and none of the listed tools would have "fully prevented" the cross-vendor pattern. They would have reduced blast radius.

View raw YAML on GitHub →