Skip to content
Lacambra Web Portal Services
← 2026, week by week
AI · Week 25, 15 to 21 June 2026

Copilot SearchLeak: the AI reads what its user can read

A one-click flaw let Microsoft 365 Copilot search everything its user could open and send it to an attacker. Microsoft fixed it; the permissions are yours.

A tall wooden library ladder on its rail against floor-to-ceiling shelves of identical plain grey box files, every shelf within its reach.

On 15 June the data security firm Varonis disclosed SearchLeak, a chain of three flaws in the enterprise search of Microsoft 365 Copilot. A victim who clicked one crafted link, on a genuine Microsoft address, set Copilot searching their mailbox, calendar, OneDrive and SharePoint, and the results went quietly to the attacker. Microsoft rated it critical, as CVE-2026-42824, and had fixed it at the start of June, with nothing for customers to do. The flaw is closed, but the lesson is not: an assistant like this can read whatever its user can read.

How one click became a search of everything

Varonis found that Copilot's search treated text in the link's address not only as a search but as instructions to follow. Those instructions told Copilot to look through the victim's mail and write what it found into the address of an image. Copilot cleaned tricks like that out of its answers, but only after the browser had started drawing them, so the image request went out first. Because the image was fetched through Bing, which the page trusted, the page's own security rules did not stop it either. The victim saw Copilot thinking for a moment, and nothing else, as BleepingComputer reported.

No permission was broken anywhere in the chain. "Copilot Enterprise operates with the user's full graph permissions," Varonis wrote, so the attacker inherited the victim's access without ever signing in: email that often holds one-time codes and password reset links, meeting notes, SharePoint documents, OneDrive files. How much an attacker could take was set by one thing, which is what that one person could open.

Permissions are the control that matters

Access in an office tends to grow. Sharing a folder with everyone takes one click, and taking it back is nobody's job, so a person's reach widens for years. Before an assistant arrives, that reach matters only when someone goes looking. An assistant goes looking in seconds, and so does anyone who can steer it.

The member and staff portals I build come down to logins, roles and permissions: what each person can see, edit and approve, enforced properly. An AI feature added to a system like that should pass the same check as the person it serves, and usually get less than that person, because a task rarely needs everything one person can open.

Storia, the wedding-planning platform I run, gives its assistant less than the couple can see. On a lawyer's advice, a guest's dietary needs never reach it, and a guest's details reach it only when the server finds a consent record. If the record is missing, the details are withheld. I wrote about that design in The AI proposes, a person confirms.

Before an assistant reads your files

  • Find what is shared with everyone: sites and folders open to the whole organisation, and links that work for anyone who has them. Close them before an assistant is switched on, not after.
  • Keep passwords, one-time codes and reset links out of email where you can. Varonis and BleepingComputer list them among what the attack could lift from a mailbox.
  • Fence off the files that matter most, such as salaries, health records and legal matters, so that access to them is a decision somebody made.
  • Treat a link that opens an AI assistant with a request already filled in as suspicious, especially a long one full of encoded characters, even on a Microsoft address.
  • Ask the supplier of any assistant what it can read, whether that can be narrowed below what its user can open, and what it records about each search.

If your own systems cannot say who can see what, that is the thing to fix before any AI goes near them, and the fit check is where to start.

Sources

  1. Varonis, SearchLeak: How We Turned M365 Copilot Into a One-Click Data Exfiltration Weapon, 15 June 2026
  2. BleepingComputer, New attack turned Microsoft 365 Copilot into 1-click data theft tool, 15 June 2026

Enrique Lacambra III builds internal business systems from Cainta, Rizal.

Email me← All writing

Tell me what is broken and I will tell you if it is worth building. A 20-minute call, no charge, and I will say so if you do not need me.