Hijacking atlassian rovo ai agent with pdf zero footprint

An innocent looking PDF document containing a hidden white text on a white background could lead to Atlassian’s Rovo AI agent leaking corporate data (such as Jira issues or Confluence documents, or even architecture diagrams) to an attacker controlled server. This security issue was discovered by cybersecurity company PromptArmor. The attack doesn’t require any user action nor does it leave any traces in the chat logs. It merely asks the AI agent to follow the hidden instruction. Although Atlassian was notified about this issue on May 23rd, they haven’t released a fix for this vulnerability within the last 2.5 months.

Enterprise AI agents with increased autonomy can introduce novel threats to organizations. The higher the level of autonomy, the easier an AI agent gets hacked. Especially, prompt injection, where hidden commands are inserted via documents, is a popular method to hack enterprise AI agents, and it may take several months from the discovery to mitigation.