DEF CON — Varonis Threat Labs has disclosed a one-click vulnerability in Rovo, Atlassian’s enterprise AI assistant, that let a specially crafted link seed attacker-controlled instructions directly into a user’s live AI session.
Dubbed RovoBlast, the flaw required no jailbreak and no permission bypass, relying on the fact that the assistant simply treated externally supplied parameters as trusted input.
Rovo functions as an AI layer spanning Jira, Confluence, Bitbucket, and third-party tools such as Slack, Microsoft 365, and Google Workspace. It also carries autonomous agent features capable of completing multi-step tasks with no further user involvement, which is what enabled the RovoBlast attack.
[ Read: How a $50,000 Exploit Chain Turned Bixby Against Samsung Phones ]
The exploit leveraged a URL parameter called rovoChatPrompt, which pre-fills content straight into Rovo’s chat window. Varonis researchers describe this attack path as parameter-to-prompt (P2P) injection, which they previously reported in Microsoft Copilot as Reprompt in January.
Researchers noticed that the organization ID part of the URL could be left blank and Atlassian would still route the request into the victim’s own default organization, all without any warning or indicator that the session had been seeded by an outside source.
To gauge the potential blast radius, the researchers simply asked Rovo what data it could see. The AI’s answer included Jira, Confluence, Bitbucket, Slack, Google Workspace, Microsoft 365, relational databases, uploaded files, web pages, and archived content.
The actual leakage came from ResearchAgent, one of Rovo’s built-in tools, which can autonomously conduct multi-source web research and navigate across arbitrary sites. Once an attacker’s prompt was seeded through the malicious link, that same capability let Rovo pull internal data and push it out to the open web in a single automated chain.
The team demonstrated the technique in three separate proof-of-concept scenarios: exfiltrating Confluence pages, Jira tickets, and SharePoint content containing personal data.
Notably, the researchers found that a single seeded link was generally enough to trigger the leak. The attack didn’t require chaining multiple requests or any additional bypass steps to get Rovo to retrieve and summarize sensitive data.
Varonis disclosed RovoBlast to Atlassian, which fixed the issue before the findings were published.
The researchers recommend that organizations limit which systems Rovo can reach, disconnect unused integrations, wall off sensitive areas such as legal, HR, and finance, disable browsing or multistep automation features that aren’t in active use, and pair this with routine monitoring of assistant activity logs.
An Atlassian spokesperson provided the following statement to SecurityWeek:
The security of our customers’ data is our highest priority. We are working with customers to implement protective controls on their instances. This is an ongoing and evolving responsibility, and we are actively working on and investing in additional solutions.
For the vulnerability to be exploited, a user with access to a customer’s Atlassian instance must provide untrusted content with a prompt injection to Rovo. This is a class of attack that affects AI systems across the industry. Similar to any phishing-type attack, we recommend customers follow security best practices and verify that any content provided to their Atlassian apps comes from a trusted source.
Varonis presented the research at DEF CON 34 on Friday. A technical write-up is available on the Varonis blog.
Related: Zero-Click AI Browser Hacking: Claude and ChatGPT Atlas Hijacked via Emails, X Posts
Related: Meta AI Hacked External Systems During Cybersecurity Testing
Related: Atlassian, Splunk Patch Critical Vulnerabilities
