Atlassian AI leaks Jira tickets via PDF hidden text

Atlassian’s AI assistant can be tricked into leaking sensitive Jira tickets and Confluence pages via hidden text buried in ordinary-looking PDFs, a security researcher warns.

Researchers at Lasso Security found that an attacker can embed invisible instructions in a PDF, telling Atlassian’s AI to copy and paste the contents of private documents into a publicly accessible URL. The AI scans the document as part of its normal workflow, processes the invisible text, and sends the data without raising any red flags. The trick works even if the PDF appears empty to human eyes.

The attack exploits how AI assistants process documents. They don’t just read what’s visible; they parse metadata, comments, and hidden layers. Atlassian’s AI assistant, part of the company’s growing suite of AI tools, isn’t designed to flag suspicious or malicious content in documents. Lasso Security demonstrated the flaw by creating a PDF with hidden text that instructed the AI to exfiltrate data to a remote server. The test was successful, showing how easily sensitive corporate information could be siphoned off.

Atlassian has not yet released a patch, but the company says it’s investigating. Security experts warn that this isn’t just an Atlassian problem; it highlights a broader risk as AI tools become more integrated into business software. Companies using AI assistants for document processing should review their security policies and consider disabling AI features that scan sensitive files until a fix is available. The incident underscores the need for AI systems to include content verification and user alerts when handling confidential data.