AI Weekly Malaysia

Back to items Summaries

Atlassian Rovo Can Be Tricked Into Sending Jira and Confluence Data to Attackers

ID
12205
Status
summarized
Published
08 Aug 2026, 4:54 PM
Fetched
08 Aug 2026, 6:08 PM
Provider
The Hacker News
Category
security
Original URL
https://thehackernews.com/2026/08/atlassian-rovo-can-be-tricked-into.html
Source URL
https://feeds.feedburner.com/TheHackersNews

Summary

Score
7.5
Created
08 Aug 2026, 6:09 PM
Tags
Audience
developersai_agent_usersai_ml_learnerssaas_founders

What happened

Two security firms independently found that Atlassian's Rovo AI assistant can be tricked into exfiltrating Jira and Confluence data to attacker-controlled servers. Varonis's 'RovoBlast' attack used the rovoChatPrompt URL parameter to preload malicious instructions and was fixed server-side on July 8, 2026. PromptArmor's attack hides instructions in uploaded files, causing Rovo to append internal data to an attacker URL with no approval step — and as of August 5, 2026 this path is not confirmed fixed.

Why it matters

If your team uses Rovo with Jira or Confluence, the file-borne prompt injection path is still live as of the report date. The only mitigation available to customers is scoping which apps and groups can access Rovo at all — not a patch. This is also a concrete example of indirect prompt injection exfiltrating real enterprise data, which is directly relevant to anyone building AI agents that read untrusted content and have tool access to sensitive systems.

Discussion angle

The PromptArmor attack works even with web-search disabled and leaves no visible trace in the user's chat — discuss how you would detect or prevent indirect prompt injection in your own AI agent architectures, and whether scoping agent access per-group is a realistic defense for Malaysian teams using Jira/Confluence.

Top