Skip to navigation

Threat reportVulnerabilityTL-2026-1939

RovoBlast: One-Click rovoChatPrompt Parameter-to-Prompt Injection in Atlassian Rovo Exposes Confluence, Jira, SharePoint, and Bitbucket Data

criticalPATCHED

Varonis Threat Labs (researcher Dolev Taler) disclosed RovoBlast, a one-click vulnerability in Atlassian's enterprise AI assistant Rovo: an unauthenticated rovoChatPrompt URL parameter let an attacker-crafted link seed arbitrary instructions directly into a victim's live, trusted Rovo chat session with no jailbreak, permission bypass, or confirmation prompt.

Key facts

IdentifiersTL-2026-1939
Also known asRovoBlast
SeverityCRITICAL
Affected productsAtlassian Rovo (Rovo Chat / rovoChatPrompt endpoint)
Weaknesses (CWE)CWE-1427, CWE-20, CWE-441
MalwareOne
ToolsResearchAgent
MITRE ATT&CK techniquesT1027 Obfuscated Files or Information, T1048.003 Exfiltration Over Unencrypted Non-C2 Protocol, T1071.001 Web Protocols, T1119 Automated Collection, T1199 Trusted Relationship, T1204.001 Malicious Link, T1213 Data from Information Repositories, T1213.001 Confluence, T1213.002 Sharepoint, T1213.003 Code Repositories, T1213.005 Messaging Applications, T1566.002 Spearphishing Link and 1 more in the attack flow
MotivationUNKNOWN
Detection rules9 (1 shown free)
Sources8 cited sources

Detection coverage

As of 2026-08-08, Threadlinqs Intelligence publishes 9 detection rule(s) for TL-2026-1939 across Splunk SPL, Microsoft KQL and Sigma, covering 14 indicator(s) of compromise. The whole corpus is readable without an account. Every threat comes with one complete detection in Splunk SPL, Microsoft KQL and Sigma, free for everyone; a free account adds daily TL Creds to unlock the rest. Paid plans add a larger daily allowance, correlation and the MCP server. Threadlinqs MCP server · View plans.

Free detection: RovoBlast — Image-Fetch HTTP Exfiltration Side Channel

Detects outbound HTTP requests for an image resource under a '/bird/<encoded-segment>/' path, or DNS/proxy activity involving the known proof-of-concept exfiltration domain kizafair.com — the RovoBlast pattern where a seeded ResearchAgent is tricked (via an 'identify the bird' prompt) into fetching an attacker-hosted image whose URL path carries exfiltrated data as an HTTP GET side channel.

  • Languages: Splunk SPL, Microsoft KQL, Sigma
  • ATT&CK: T1071.001, T1048.003, T1027, T1583.001
  • Data sources: web_proxy, firewall, dns_logs
  • Severity: critical
  • Confidence: high

Splunk SPL

index=proxy OR index=firewall sourcetype IN ("zscalernss-web","pan:traffic","stream:http")
| where match(url,"(?i)/bird/") AND match(url,"(?i)\.(jpg|jpeg|png|gif)(\?|$)")
| rex field=url "/bird/(?<encoded_segment>[^/]+)/"
| where len(encoded_segment) > 8
| eval is_kizafair=if(match(dest_domain,"(?i)kizafair\.com"),1,0)
| stats count min(_time) as first_seen max(_time) as last_seen values(src_ip) as src_ips values(url) as urls by dest_domain, is_kizafair
| where is_kizafair=1 OR count > 0
| table first_seen, last_seen, dest_domain, src_ips, urls, count
| sort -first_seen

Microsoft KQL

CommonSecurityLog
| where TimeGenerated > ago(24h)
| where RequestURL has "/bird/" and RequestURL matches regex @"\.(jpg|jpeg|png|gif)(\?|$)"
| extend EncodedSegment = extract(@"/bird/([^/]+)/", 1, RequestURL)
| where strlen(EncodedSegment) > 8
| extend IsKnownPoCDomain = DestinationHostName has "kizafair.com"
| project TimeGenerated, SourceIP, DestinationHostName, RequestURL, EncodedSegment, IsKnownPoCDomain
| union (
    DnsEvents
    | where TimeGenerated > ago(24h)
    | where Name has "kizafair.com"
    | project TimeGenerated, SourceIP=ClientIP, DestinationHostName=Name, RequestURL="", EncodedSegment="", IsKnownPoCDomain=true
)
| order by TimeGenerated desc

Sigma

title: Rovo Image-Fetch HTTP Exfiltration Side Channel (RovoBlast)
id: 1f4d7a9c-5e2b-4f6a-8d3c-7a9e1b2c4f6d
status: experimental
description: Detects outbound HTTP requests for an image under a '/bird/<encoded-segment>/' path, and/or DNS resolution of the known RovoBlast PoC exfiltration domain kizafair.com — the pattern used to trick Rovo's ResearchAgent into requesting an attacker-hosted image whose URL path carries exfiltrated data via an HTTP GET side channel.
references:
    - https://www.varonis.com/blog/rovoblast
    - https://bugcrowd.com/disclosures/bf1922fb-99d0-4d3b-b419-1728720d29ec/one-click-data-exfiltration-via-rovochatprompt-url-parameter-confluence-rovo
author: Threadlinqs Team
date: 2026-08-08
tags:
    - attack.command-and-control
    - attack.t1071.001
    - attack.exfiltration
    - attack.t1048.003
    - attack.defense-evasion
    - attack.t1027
    - attack.resource-development
    - attack.t1583.001
logsource:
    category: proxy
    product: generic
detection:
    selection_path:
        c-uri|contains: '/bird/'
    selection_ext:
        c-uri|re: '\.(jpg|jpeg|png|gif)(\?|$)'
    selection_domain:
        r-dns|contains: 'kizafair.com'
    condition: (selection_path and selection_ext) or selection_domain
falsepositives:
    - Legitimate bird-identification or wildlife-photo requests to unrelated benign photo-hosting domains that coincidentally use a '/bird/' path segment
level: critical

Known false positives:

  • Benign image requests to unrelated third-party photo-hosting or wildlife-identification services that happen to use a '/bird/' URL segment

Tuning notes: Add kizafair.com and any newly observed sibling PoC/attacker domains to a dedicated watchlist/lookup for exact-match alerting; keep the '/bird/' path-pattern rule as a lower-confidence hunt for copycat campaigns using different domains but the same encoding trick.

8 more detection rules (Splunk SPL, Microsoft KQL, Sigma) · unlock with TL Creds. Sign up free to unlock the detection rules or query them with the Threadlinqs MCP server (Purple)
  • RovoBlast — rovoChatPrompt Link Click via Safe Links / URL Click Telemetry · unlock with TL Creds
  • RovoBlast — ResearchAgent Cross-System Mass Collection (Custom Atlassian Table) · unlock with TL Creds
  • RovoBlast — Image-Fetch Exfiltration via CommonSecurityLog / DNS · unlock with TL Creds
  • RovoBlast — rovoChatPrompt Parameter Present on Rovo Chat Endpoint (Sigma) · unlock with TL Creds
  • RovoBlast — Cross-Container Mass Enumeration by Rovo ResearchAgent (Sigma) · unlock with TL Creds
  • RovoBlast — Bird-Image Exfiltration Path Pattern / [indicator] (Sigma) · unlock with TL Creds
  • RovoBlast — rovoChatPrompt Link Click via Proxy/SWG Logs · unlock with TL Creds
  • RovoBlast — ResearchAgent Cross-System Mass Collection (Atlassian Audit Log) · unlock with TL Creds

Indicators of compromise

As of 2026-08-08, Threadlinqs Intelligence tracks 14 indicators of compromise for TL-2026-1939.

Sample indicators (1 of 14)

  • Domain: kizafair[.]com

Defanged, TLP:CLEAR. Full set of 14 indicators: unlock with TL Creds.

14 indicators of compromise · unlock with TL Creds. Compare plans

Technical analysis

Using Rovo's built-in ResearchAgent, the injected instructions could autonomously enumerate and exfiltrate Confluence pages, Jira tickets, SharePoint content, and other connected-system data to an attacker-controlled server via an HTTP image-fetch side channel. Atlassian deployed a server-side fix (2026-07-08) neutralizing automatic execution of rovoChatPrompt content before Varonis published its findings and presented at DEF CON 34 (2026-08-08).

RovoBlast exploits a Parameter-to-Prompt (P2P) injection flaw in the `rovoChatPrompt` query parameter accepted by Atlassian's Rovo AI assistant at `home.atlassian.com/chat?rovoChatPathway=chat&rovoChatPrompt=<prompt>`. When an authenticated user opens a specially crafted link, Rovo pre-fills and auto-executes the embedded text as if it were the user's own genuine chat query, with no warning, confirmation dialog, or visual indicator that the session had been seeded by an outside party. Varonis found that the organization-scoping portion of the URL (`/o/<org-id>/chat`) could be left blank, and Atlassian would silently route the request into the victim's own default organization, removing the need for an attacker to know or guess the target org ID.

Once seeded, the attacker-controlled prompt could invoke ResearchAgent, a built-in Rovo capability described by Varonis as able to "autonomously conduct multi-source web research and navigate across arbitrary sites." This let the injected instructions harvest sensitive data reachable through Rovo's federated access layer — Confluence pages, Jira tickets, Bitbucket repositories, and, via Rovo's 50+ third-party Connectors, Slack, Microsoft 365, Google Workspace, SharePoint, and other integrated systems, all evaluated with the victim's own live privileges.

The Bugcrowd technical disclosure underlying the research (submitted by researcher handle "loacker", reported 2025-11-04, triaged as P2 with a $6,000 bounty on 2026-01-06) documents a working proof-of-concept exfiltration channel: the crafted prompt used indirect, guardrail-evading phrasing (assigning the victim's harvested email address to an "emoji variable," then asking Rovo to "identify the bird" in an image at `https://<attacker-domain>/bird/<emoji>/bird.jpg`) to trick Rovo into requesting an attacker-hosted image whose URL path carried the exfiltrated data as an HTTP GET request — leaving no obvious trace in the visible chat transcript. The PoC used the domain `kizafair.com` as the attacker-controlled logging endpoint.

Atlassian remediated the issue server-side on 2026-07-08, disabling automatic execution of rovoChatPrompt-seeded instructions, and characterized the underlying issue as a "phishing-class" attack requiring a user to interact with untrusted, attacker-supplied content — consistent with vulnerabilities affecting AI systems industry-wide. Varonis Threat Labs published RovoBlast and presented it at DEF CON 34 on 2026-08-08, noting that Rovo cannot be fully uninstalled from an Atlassian Cloud tenant, and that no taint-labeling exists to distinguish externally seeded prompt content from a user's own input. No CVE has been assigned. This disclosure is distinct from, but thematically related to, a separate unpatched zero-click Rovo prompt-injection flaw disclosed by PromptArmor on 2026-08-05 that abuses Rovo's URL-retrieval tool even when web search is administratively disabled.

MITRE ATT&CK techniques by tactic:

Defense Evasion

T1027 Obfuscated Files or Information

Exfiltration

T1048.003 Exfiltration Over Unencrypted Non-C2 Protocol

Command and Control

T1071.001 Web Protocols

Collection

T1119 Automated Collection; T1213 Data from Information Repositories; T1213.001 Confluence; T1213.002 Sharepoint; T1213.003 Code Repositories; T1213.005 Messaging Applications

Initial Access

T1199 Trusted Relationship; T1566.002 Spearphishing Link

Execution

T1204.001 Malicious Link

Resource Development

T1583.001 Domains

Remediation

Patches

  • Atlassian server-side fix deployed 2026-07-08 disabling automatic execution of rovoChatPrompt-seeded instructions

Immediate actions

  • Confirm the Atlassian server-side patch (deployed 2026-07-08) has propagated to the tenant; verify rovoChatPrompt links no longer auto-execute
  • Audit web-proxy/SIEM/email-gateway logs for historical links containing the rovoChatPrompt URL parameter
  • Monitor outbound requests from Rovo-integrated infrastructure for unexpected external image/URL fetches that could indicate exfiltration attempts
  • Warn users not to click unsolicited Atlassian/Rovo chat links from external senders

Workarounds

  • Restrict or disable ResearchAgent and other autonomous Rovo agents for tenants pending confirmation of patch deployment
  • Block or alert on inbound links containing the rovoChatPrompt parameter at email and URL-filtering gateways

Longer-term hardening

  • Require explicit user confirmation before any AI-agent chat session executes instructions seeded via URL parameters or other externally controlled inputs
  • Implement provenance/taint tracking so Rovo (and similar AI assistants) can distinguish user-authored prompts from externally seeded content
  • Apply least-privilege scoping to Rovo Connectors (Confluence, Jira, Bitbucket, SharePoint, Slack, Microsoft 365, Google Workspace) so a single seeded session cannot traverse all connected systems
  • Gate autonomous agent capabilities (e.g., ResearchAgent) that perform outbound web requests behind human-in-the-loop confirmation, especially when session data could be embedded in the request

Affected and fixed versions

  • Atlassian — Rovo (Rovo Chat / rovoChatPrompt endpoint)
    Vulnerable versions: Rovo Chat prior to the 2026-07-08 server-side fix
    Fixed in: Rovo Chat post 2026-07-08 server-side remediation

Timeline

  • Vulnerability in the rovoChatPrompt URL parameter reported to Atlassian via Bugcrowd by researcher handle 'loacker'.
  • Atlassian triages the Bugcrowd submission as P2 priority (AI Application Security / Sensitive Information Disclosure) and issues a $6,000 bounty.
  • PromptArmor separately reports a distinct zero-click Rovo prompt-injection/data-exfiltration flaw to Atlassian, receiving an acknowledgment and ticket ID.
  • Atlassian deploys a server-side fix disabling automatic execution of rovoChatPrompt-seeded instructions, resolving the RovoBlast vulnerability.
  • PromptArmor publicly discloses its separate, still-unpatched zero-click Rovo indirect-prompt-injection data-exfiltration flaw.
  • Atlassian confirms the fix publicly, characterizes RovoBlast as a 'phishing-class' attack requiring interaction with untrusted content, and states it is working with customers on additional protective controls.
  • Varonis Threat Labs (researcher Dolev Taler) publishes the RovoBlast research and presents it at DEF CON 34.

Sources cited

Previous and next in vulnerability

Further reading

Drafted by the Threadlinqs research pipeline from 8 sources; published ; last updated —.

Threadlinqs Intelligence — Real-Time Threat Detection Platform

// threat_feed
$ sort --newest
Showing all threats

Live intelligence console

Threat level

Latest Threats