Microsoft Copilot Cowork Exfiltrates Files

This attack achieved a high success rate against state-of-the-art models, including Claude Opus 4.7.

Overview

Copilot Cowork is a Frontier feature available now in Microsoft 365. It operates with the users’ Microsoft permissions and can use Microsoft Graph to read and operate on data in one’s Microsoft tenant.

In this article, we demonstrate that through an indirect prompt injection in a poisoned skill, attackers can exfiltrate files from M365. This is done by exploiting the fact that, unlike other sensitive actions, sending emails and Teams messages to the active user does not require human approval, and opening the compromised messages in Teams or Outlook can trigger attacker-controlled network requests.

This risk reflects that giving agents access to multiple systems expands the prompt-injection attack surface. In isolation, the agent’s intended capabilities are benign; however, due to the properties of the integrated systems, users are at risk. This is reminiscent of our previous work on how URL previews in communications apps have become an egress surface for agents. As this risk pertains to the design of a system in which agents act with delegated authority across an entire enterprise ecosystem, rather than to a specific bug, we are publicizing this work to inform users of the risks they are accepting by using an agentic product of this nature.

Separate from this risk, we have disclosed a vulnerability to Microsoft that directly allows data egress from Copilot Cowork’s sandbox environment.

The Attack Chain

Microsoft’s documentation on action approvals states, “[Copilot] Cowork asks for your permission before taking sensitive actions, like sending an email or posting a message in Teams.” However, in practice, when the recipient is the active user, these actions execute immediately without requiring human approval (users do not have a setting to modify this behavior). Because these messages can contain external images that trigger network requests to external websites, data can be exfiltrated when a user opens a compromised message sent by the agent. Copilot Cowork can retrieve ‘pre-authenticated download links’ for files the user has access to, which allow anyone who opens the link to download that file. So, a manipulated agent can exfiltrate files by exfiltrating pre-authenticated download links.

1. The victim has access to files stored in SharePoint or OneDrive containing PII & Financial data

   

2. The victim uploads a skill file to Copilot Cowork that contains a prompt injection

   For general use cases, this is quite common; a user finds a file online that they upload as a skill. This attack is not dependent on the injection source - other injection sources include, but are not limited to: web data from Claude for Chrome, connected MCP servers, etc.

   

   Note: Admins have limited oversight of ‘Skills’, as Skills in Copilot Cowork are automatically loaded from a specific path in a user’s OneDrive.

   
   

3. The victim asks Microsoft Copilot Cowork to review what they worked on that week, triggering the skill

   

   
   

4. The injection manipulates Microsoft Copilot Cowork to post a Teams message that will exfiltrate pre-authenticated file download links when it is viewed

   The injection tells Copilot Cowork that a service exists to create document previews for the recap message; to do this, the agent retrieves pre-authenticated file download links for each file and passes those URLs as query parameters to an attacker-controlled site via malicious HTML image tags. 

   

   At no point in this process is human approval required.

   If we expand the ‘Task complete’ block, we can see the agent’s actions play out – but the malicious message content is never visible, even when the Teams action is clicked on.

   

5. When the user opens their Teams messages, the pre-authenticated download links are exfiltrated, and the attacker can download the files by visiting the link

   

Mitigating Risks for Your Organization 

Microsoft Copilot Cowork has read access to essentially any resource a user does through Microsoft Graph. As such, the primary mechanism to reduce the blast radius of attacks like this is to restrict excessive permissioning across one’s Microsoft ecosystem.

To restrict users’ ability to retrieve pre-authenticated download links for files, administrators can restrict file downloads from SharePoint by running commands in the SharePoint Online Management Shell:

Set-SPOSite -Identity <SiteURL> -BlockDownloadPolicy $true

Or, to block based on a sensitivity label:

Set-Label -Identity <label> -AdvancedSettings @{BlockDownloadPolicy="true"}

Note: This configuration affects functionality; documentation states that for files under the policy 'BlockDownloadPolicy', "Users have browser-only access with no ability to download, print, or sync files. They also can't access content through apps, including the Microsoft 365 Apps (like Word, Excel, PowerPoint, and so on)."

Model Agnostic Exploitation

The attack chain was initially conducted with the model selection set to ‘auto’, which dynamically routes between Claude Opus 4.7 and Claude Sonnet 4.6. However, we validated explicitly that this attack succeeds with the exact same injection on the more advanced Opus 4.7 model by setting the model directly.

Opus 4.7 was more comprehensive in its search for recently edited documents; it expanded exfiltration to include every document used in previous Cowork Copilot sessions that week, as well as the files stored in more typical document locations that were found when the model was set to ‘Auto’.

Prompt Injection Efficacy

This prompt injection exhibited a very high efficacy, and we noted that Copilot Cowork completed the entire attack chain on every trial (5 for 5). Furthermore, the attack was not contingent on the specific wording of the user query – whenever the model invoked the skill, the injection succeeded.

The injection consisted of 5 lines in an 81-line skill file, all of comparable length to the other lines.

This demonstrates that even with the latest models and only a small excerpt of malicious text, an indirect prompt injection can hijack agent behavior. 

As such, we urge readers to exercise caution when working with untrusted data, such as skills shared online – especially when the untrusted data is placed into a trusted context, such as a skill file.