Cybersecurity Researchers Uncover Flaw in Google AI Coding Tool

Cybersecurity researchers from Pillar Security discovered a flaw in Google's AI coding tool that could allow attackers to gain Editor-level access to an internal Google Cloud project. The vulnerability was found in the command line interface (CLI) setup code of Google's AI tool, Gemini. A researcher from Pillar Security was able to exploit this flaw by filing a "bug report" containing hidden instructions that resulted in a prompt injection.

This injection led to a legitimate credentials file being issued via the Workload Identity Federation (WIF) framework, which the researcher then copied out. One of these credentials allowed the researcher to impersonate a more powerful account, granting them Editor-level control over the project. This breach highlights the growing risk of compromising software supply chains using AI technologies.

While this incident is isolated, it demonstrates how easy it is to exploit open source tools in the AI coding era. Cybercriminals could potentially compromise an entire software supply chain at machine speed by inserting malicious prompts into coding agents. DevSecOps teams should closely monitor the data sources an AI coding tool accesses to mitigate this risk.

This is not the first time Google's AI technologies have been found vulnerable by Pillar Security. Earlier this month, the researchers revealed the first instance of an AI agent that could exploit another AI agent, posing a significant security concern. Google has since addressed these issues, but the potential for new vulnerabilities remains high given the rapid development of AI technologies.