Latest / Decrypted: The UK Cyber Briefing / Atlassian's Rovo assistant, and the data leak it hasn't fully fixed
Transcript
- 0:03This is Decrypted. Saturday, the eighth of August.
- 0:06Atlassian's Rovo assistant, and the data leak it hasn't fully fixed.
- 0:12Atlassian's Rovo is the AI assistant now built into Jira and Confluence, and it's switched on by default for
- 0:20most customers. This week it failed two independent break-in attempts.
- 0:25Both attackers were security researchers.
- 0:28Both attacks worked. And one of them still isn't fixed.
- 0:32Varonis and PromptArmor each found a way to get Rovo to read a tenant's own tickets and documents, then quie
- 0:40tly post the contents to an address of the attacker's choosing.
- 0:43No stolen password. No malware.
- 0:46Just a document or a link with an instruction hidden inside it.
- 0:50Atlassian is used right across UK organisations, from universities to N H S trusts to central government.
- 0:58So this one is worth understanding, rather than skimming past.
- 1:02Next. Two ways in, one still open.
- 1:06Varonis calls their version RovoBlast.
- 1:10It used a crafted link. Rovo's chat interface accepts a U R L parameter that pre-loads a prompt into a sessio
- 1:18n, and it does that without asking the user to confirm anything.
- 1:22One click, and Rovo's autonomous research agent starts pulling Confluence pages, Jira tickets and connected S
- 1:30harePoint content. Then it pushes it all out through an ordinary image request to an external server.
- 1:36Atlassian fixed this server-side in July, after Varonis reported it through Bugcrowd.
- 1:42The pair presented the research at D E F C O N 34.
- 1:46PromptArmor's version doesn't need a link at all.
- 1:49Hide an instruction inside a document that Rovo will legitimately be asked to read.
- 1:54The assistant then builds a U R L with the sensitive data added on the end, and fetches that U R L itself to
- 2:01check it. The data leaks in the act of retrieval.
- 2:05It works even when an administrator has switched off Rovo's web search, because the underlying fetch capabili
- 2:11ty is separate from search.
- 2:13PromptArmor reported this to Atlassian on 23 May.
- 2:17After an initial acknowledgement, Atlassian went quiet.
- 2:20And as of PromptArmor's public disclosure on 5 August, the flaw was still live.
- 2:26And finally. The design fault sits underneath both.
- 2:29Both attacks work for the same reason.
- 2:32Rovo has no boundary between content it reads and actions it takes.
- 2:37An instruction buried in a Confluence page or a U R L parameter is treated with the same trust as a request t
- 2:45yped by the logged-in user.
- 2:46And nothing checks whether a web address the agent constructed itself is a safe place to send data before it
- 2:53goes there.
- 3:01oes next. S Q L injection taught the industry to stop trusting user input to build queries.
- 3:07The same discipline now has to apply to anything an AI agent reads before it acts on it.
- 3:12For UK organisations, the practical question isn't "do we use Rovo's chat".
- 3:18It's "what have we already switched on".
- 3:21Enterprise AI assistants increasingly ship enabled by default across a SaaS estate, with connectors to Jira,
- 3:29Confluence, SharePoint and Outlook granted the moment a licence is turned on.
- 3:33So, worth doing this week.
- 3:36Check whether Rovo, Copilot, or an equivalent is live in your tenant.
- 3:41Confirm what it's allowed to read.
- 3:43And ask your vendor whether outbound requests the agent makes on its own are restricted to an allowlist, rath
- 3:50er than left open to anywhere on the internet.
- 3:52That was Decrypted. The written version and every source are linked in the show notes.
- 3:58Thanks for listening.