Threat Intelligence

GitHub Issues Used to Hijack AI Coding Assistants' CI Pipelines

The Hacker News · 7 Aug 2026

Key Takeaway If your business uses AI coding assistants connected to CI/CD pipelines, review their default permissions and restrict what external, unauthenticated inputs like GitHub issues can trigger.

Researchers at Novee Security have revealed a serious flaw affecting popular AI coding assistants, including Anthropic's Claude Code, Google's Gemini CLI, and OpenAI's coding agent. In their default configurations, these tools could be manipulated through something as simple as a GitHub issue submitted by an outsider with no special access to the repository.

The researchers demonstrated that opening a GitHub issue was enough to trigger code execution on the continuous integration (CI) runners behind Anthropic's and Google's own repositories. For OpenAI's agent, the same technique allowed attackers to hijack the next automated run of the coding agent. This means an attacker with zero privileges could potentially influence sensitive automated workflows and access secrets stored within CI environments, such as API keys or credentials.

The findings, presented at Black Hat USA on August 5, highlight a growing risk as businesses increasingly adopt AI-powered coding tools that interact automatically with code repositories and CI/CD pipelines. Because these agents are designed to respond to inputs like issues and comments, attackers can exploit that automation to reach systems that were assumed to be protected by repository permissions.

AI Security CI/CD Security GitHub Supply Chain Risk Coding Assistants

Summarised by CISO AI from The Hacker News. We link back to every original so you can read it yourself.