'GitLost' Flaw Exposes Private Data from GitHub's Agentic Workflows

A critical vulnerability named 'GitLost' has been discovered in GitHub's Agentic Workflows, allowing unauthenticated attackers to exploit public GitHub Issues to access private repository data. The flaw enables attackers to craft an issue in a public repository, which the AI agent then processes, inadvertently leaking sensitive information from private repositories. Researchers from Noma Security identified this prompt injection vulnerability, emphasizing that it does not require any credentials or coding skills to exploit. The AI agent's context window serves as both a functional tool and an attack surface, leading to potential data exposure. The proof of concept (PoC) demonstrated how an attacker could hide commands within the issue body, which the AI agent would execute. This vulnerability raises significant concerns regarding the security of AI-driven automation in software development. Organizations using GitHub's Agentic Workflows are urged to review their configurations and implement security measures to mitigate this risk.

Key Points: • The 'GitLost' vulnerability allows unauthenticated attackers to leak private data from GitHub. • Attackers can exploit public GitHub Issues without needing credentials or coding skills. • The flaw highlights significant security challenges in AI-powered automation systems.