Vulnerability Found in Atlassian's AI Agent 'Rovo': Risks of Sensitive Data Extraction via Hidden PDF Text

Overview of the Vulnerability

A serious security risk has been identified in Atlassian’s AI agent, Rovo. Investigations have revealed that the AI can parse text intentionally hidden within PDF files, leading to the potential for unauthorized extraction and output of sensitive information.

Mechanism and Background

This issue stems from the document analysis process used by Rovo, an AI agent designed to leverage organizational knowledge across various platforms. When utilizing RAG (Retrieval-Augmented Generation) technology, AI systems may reference not only the visible text within a document but also metadata or text with hidden attributes. By exploiting this behavior, attackers can force the AI to read sensitive information that is visually obscured, creating a risk of unauthorized data access.

Security Challenges in AI Development

This incident highlights the critical importance of input data reliability and the design of security boundaries in business tools powered by Large Language Models (LLMs). The fact that hidden text within a PDF can function as a "data leakage vector" to bypass user security boundaries represents a new blind spot for AI product developers.

Future Measures and Expectations

AI vendors, including Atlassian, are now under pressure to enhance pre-processing and filtering capabilities within their AI data pipelines. Implementing strict access controls for invisible information within documents and redesigning output guardrails for AI agents will become increasingly essential development requirements moving forward.