Incident · Jul 2026 · Prompt injection

Atlassian Rovo

Security researchers showed that crafted content could steer Rovo into following attacker-controlled instructions and exposing information available to the signed-in user. A separate report showed that hidden instructions in workspace content could make the assistant send Jira and Confluence material outward even with web search turned off.

What happened

Security researchers showed that crafted content could steer Rovo into following attacker-controlled instructions and exposing information available to the signed-in user. A separate report showed that hidden instructions in workspace content could make the assistant send Jira and Confluence material outward even with web search turned off.

The weakness started with content that Rovo treated as instructions instead of as untrusted data. A crafted link or embedded prompt could be interpreted by the assistant during normal use.

Because the assistant was allowed to act on retrieved workspace content, the injected instructions could redirect it toward data the user could already access. That made the abuse look like normal assistant behavior rather than a blocked attack.

The affected material included workspace data from Jira and Confluence, and in one case data could be forwarded to an outside server. Atlassian fixed the link-based issue on July 8, 2026, while the content-based path was still reported as unresolved when published later.

A stronger boundary between retrieved content and executable instructions, plus strict output controls on external transmission, would have stopped the abuse earlier.

Source: Vibe Graveyard ↗. Follow it for the full account.

The cause: prompt injection

Text the assistant reads (an email, a web page, a file name) carries instructions, and the assistant follows them as if they came from you.

Framework entries that cover it: OWASP LLM01: Prompt Injection · MITRE ATLAS: LLM Prompt Injection

The solution
  1. 1

    What would have caught this one

    A check that labels all retrieved page text and links as untrusted input, then blocks any attempt to turn that text into instructions or outbound exfiltration, would have caught it.

  2. 2

    Prevent the whole class

    Treat everything the assistant reads as untrusted. Separate reading from acting, and require confirmation before any action that sends data out.

  3. 3

    Test for it

    Plant instructions in every field the assistant can read, including encoded and split ones, and check whether any action or data leaves.