On 10 September, Huntress started seeing phishing that pointed at app.powerbi.com 1. The link led to Microsoft’s own service, and the page behind it was a real Power BI report the attacker built to look like a blurred document with a Download Reference button. A click on that button opened a new tab on a domain the attackers controlled, and the script on that page sent a ScreenConnect installer only after it had profiled the visitor. The first thing I check on a link in mail is what the host has done lately. I ran that check on app.powerbi.com and found nothing to hold the message.
The outer link is genuine
An attacker with an account builds a report on app.powerbi.com, puts a link inside it, sets the report to public, and mails the URL 1. The URL carries a base64 parameter that decodes to a report id and a tenant id. On the lure Huntress published, the report id is 5946eb49-c35c-4210-92e7-4e9e2fc33b10 and the tenant id is 55a28be6-1bc4-4231-9104-7d2aea5f0fa6 1. The report behind the link is real and Microsoft hosts it 1. A gateway that trusts the domain has nothing to weigh, and Huntress describes the effect as a link that “skirts through Microsoft 365 mail filters and other security gateways that trust this domain” 1.
Decoding it does not settle much. On most days a valid report on a valid tenant is a business intelligence dashboard someone shared, and the parameter alone does not make it suspicious.
Fingerprinting before the payload
Clicking Download Reference opened a new tab to dailylifeproject.site in the incident Huntress documented, with burnsworth.site, essaywritingservice.site and openpediatrics.site in other incidents 1. Before anything downloaded, the page collected the visitor’s operating system, browser and version, whether the session was mobile or desktop, the user agent, automation indicators, screen size, iframe context, and cloud-provider cookies 1. It carried a Telegram bot token and chat id, and sent the visitor’s IP address, geolocation, browser, operating system and download activity to that bot 1. A visitor who failed the checks went to a redirect page instead 1. One variant allowed only Windows desktops and filtered traffic from Microsoft or unknown ISPs 1.
My detonation fits the category that page screens out: a datacenter address, an automated browser, no consumer cookie history. If the page reads those signals, the verdict I get describes a redirect page. The payload stays behind it. I got a clean score on that URL because the visitor the page wanted never arrived.
A few seconds is enough
The download is delayed on purpose. After a few seconds, a script clicks a hidden link, and the page displays a notification that the “Reference Verification Form downloaded successfully” 1. A detonation that stops at the first render records a page with no file. A person who stays on the page gets the installer 1. The delay costs the attacker nothing and breaks my assumption that the first request carries the goods.
So when a link points at a trusted cloud service and the page asks for a download, I treat the detonation verdict as a fact about one visitor’s session. It describes the page that address and browser received.
Two clients, and a task that puts one back
The installer came from a ScreenConnect tenant at hamham27.screenconnect.com, with the t guest-access parameter changing across campaign variants: ILEAYEASAN, PERFECTO and PAPASUPE 1. Huntress says that change suggests campaign- or lure-specific access tokens 1. The first client, 2b302081e9e777d0, connected to instance-g01s1n-relay.screenconnect.com, then pulled a second client, 43773b3da4ccb17b, pointed at onthegotree.site 1. In one incident a CMD file launched SCAutoFix.ps1, which downloaded the second installer and uninstalled the first client 1. Once the rogue clients were in place, HideUL_x64.exe ran, and Huntress assesses it as a tool for hiding activity from the user and from security software 1. Huntress also records one incident where a scheduled task named SCAutoRepairEvery2Min ran SCAutoFix.ps1 every two minutes 1. A retroactive hunt tied the same client and configuration to 22 other endpoints in separate incidents 1.
Removal has to count every client. A responder who removes only the client it can see and stops there leaves the second one running, and the scheduled task launches SCAutoFix.ps1 again within two minutes, so the script re-downloads and runs the additional installer 1. The order I send is the task first, then both client services, then the installers and the script, then a check that the relay domains are blocked. Anything that treats this as one uninstall leaves a working remote access channel behind.
What I check now
The outer URL still tells me something. The report id and tenant id show whether the dashboard belongs to us. A report under a tenant I do not recognize is a link to somebody else’s dashboard, and a mail message has no reason to carry one. That check costs a decode.
The button sits inside the report definition. I can wrap the app.powerbi.com URL when I rewrite links. The visitor’s browser builds the link behind the button as the report opens, which is after my rewrite has run. When I block a lure, I block one report id in one tenant. The attacker stands up another tenant, and I build the block list again.
On the endpoint side I look for the client service by name and install path, the SCAutoRepairEvery2Min task, runs of ScreenConnect.ClientSetup.exe from a Downloads folder, and outbound connections to ScreenConnect relay hosts that are not on our approved list. The two-client design means one hit is not the whole count, and I give an endpoint that looks clean after a single uninstall another look.
The user-facing tell is a document that needs an install before it opens. The attacker dresses it as a blurred reference document, and the visitor who clicks Download Reference ends up with an executable. Training that names that shape does more than a warning about unknown senders, because the link points at Microsoft’s real Power BI domain 1.
The tenant setting
Microsoft’s documentation states that Publish to web requires a Power BI administrator to enable the tenant setting, and that the administrator can scope it to the entire organization or to specific security groups 2. If my organization does not publish dashboards to the open web, that setting has no reason to be on, and I can ask for it off or scoped to a small group.
Paubox notes that the setting does not stop an outside attacker using their own account, and that it does stop a compromised staff account from hosting lures on the organization’s own tenant 3. The second half is the part the tenant owner can fix.
The campaign started in September and the reporting is a month old. I keep it for the mechanics. I read who owns the domain off the hostname, and took everything past that from the fingerprinting page, the delay and the second client.
-
Reddish, S. (2026, October 7). Phishing campaign abuses Microsoft Power BI to deploy rogue RMMs. Huntress. https://www.huntress.com/blog/screenconnect-power-bi ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎ ↩︎
-
Microsoft. (2025, November 20). Publish to web from Power BI. Microsoft Learn. https://learn.microsoft.com/en-us/power-bi/collaborate-share/service-publish-to-web ↩︎
-
Amod, F. (2026, November 2). Phishing campaign hides lures inside public Power BI dashboards. Paubox. https://www.paubox.com/blog/phishing-campaign-hides-lures-inside-public-power-bi-dashboards ↩︎