Security firm PromptArmor has shown how hidden text inside a PDF can be used to hijack Atlassian’s Rovo AI agent, according to The Decoder. The reported attack can cause the agent to forward sensitive information from Jira and Confluence to an external server.
The case is an example of prompt injection, where malicious instructions are placed in content an AI system reads. Because agents can connect to business tools and take actions, the risk is greater than a misleading chatbot response: a compromised agent may access or move data.
The practical lesson is that enterprise AI agents need defenses around documents, permissions, and outbound actions. Hidden instructions should not be treated as trusted commands, and sensitive operations should require strong controls. The report also shows why organizations cannot secure agents only by improving the model; they need system-level safeguards around what the agent can read and do.