Microsoft Copilot Vulnerable to Indirect Prompt Injection, Hidden Report Reveals
Safety research firm PromptArmor recently published a report highlighting a serious security flaw in Microsoft's AI agent service Copilot Cowork, which is part of Microsoft 365. Attackers can use a method known as "indirect prompt injection" to quietly steal and leak confidential files from an organization's internal cloud storage without user consent.

Malicious Instructions Concealed in Office Templates
As an integrated AI assistant, Cowork has broad permissions to send emails, post messages in Teams, and access internal data from OneDrive and SharePoint. However, researchers have discovered that attackers can embed harmful instructions within web pages, documents, or seemingly routine office automation templates, such as a "Weekly Work Review," to trick the AI agent into executing them.
When a user asks Cowork to handle a file that contains these malicious prompts, the AI agent is deceived into falsely claiming it needs to generate a document preview. Then it automatically retrieves pre-authenticated download links for sensitive files and sends those links back to the attacker through Teams messages—all carried out in the background, making it extremely hard for users to notice.
Scheduled Tasks Amplify Risk and Are Hard to Block
The report notes that because Copilot Cowork can perform tasks on a scheduled basis, this significantly increases the security danger. For instance, recurring automated tasks like "Weekly Report Summary," which run periodically, can repeatedly trigger and execute the attack chain in the background—even when the user is away from their screen and not actively using the system.
In security tests, this attack method achieved a 100% success rate across five trials. Worse still, administrators have limited ability to monitor or control such "skill files," and the vulnerability is not only effective in automatic mode but also remains exploitable when explicitly calling more powerful large models like Claude Opus 4.7.
Related article
Lei Jun confirms Xiaomi's desktop AI agent MiClaw in development, MiMo-V2-Pro launches across all platforms
At the 2026 China Development High-level Forum, Xiaomi Group's Lei Jun confirmed that the long-awaited desktop version of the AI agent "MiClaw" (crab) is now on the development roadmap. Xiaomi had already launched a limited closed beta for the mobile
OpenAI Restarts Robot Business, Automan Seeks Engineers for Infrastructure R&D
On June 1st, OpenAI CEO Sam Altman announced on social media that the company is re-entering the robotics field, releasing job openings for the OpenAI Robotics team. The company is hiring full-stack hardware, operations, systems, and machine learning
Bain forecasts US$100 billion SaaS market in agentic AI automation
Bain & Company has estimated a $100 billion market in the U.S. for SaaS companies leveraging agentic AI. The firm said this market stems from automating coordination tasks within enterprise systems.This estimate comes from the second installment in B
Related Special Topic Recommendations
Comments (0)
0/500
Safety research firm PromptArmor recently published a report highlighting a serious security flaw in Microsoft's AI agent service Copilot Cowork, which is part of Microsoft 365. Attackers can use a method known as "indirect prompt injection" to quietly steal and leak confidential files from an organization's internal cloud storage without user consent.

Malicious Instructions Concealed in Office Templates
As an integrated AI assistant, Cowork has broad permissions to send emails, post messages in Teams, and access internal data from OneDrive and SharePoint. However, researchers have discovered that attackers can embed harmful instructions within web pages, documents, or seemingly routine office automation templates, such as a "Weekly Work Review," to trick the AI agent into executing them.
When a user asks Cowork to handle a file that contains these malicious prompts, the AI agent is deceived into falsely claiming it needs to generate a document preview. Then it automatically retrieves pre-authenticated download links for sensitive files and sends those links back to the attacker through Teams messages—all carried out in the background, making it extremely hard for users to notice.
Scheduled Tasks Amplify Risk and Are Hard to Block
The report notes that because Copilot Cowork can perform tasks on a scheduled basis, this significantly increases the security danger. For instance, recurring automated tasks like "Weekly Report Summary," which run periodically, can repeatedly trigger and execute the attack chain in the background—even when the user is away from their screen and not actively using the system.
In security tests, this attack method achieved a 100% success rate across five trials. Worse still, administrators have limited ability to monitor or control such "skill files," and the vulnerability is not only effective in automatic mode but also remains exploitable when explicitly calling more powerful large models like Claude Opus 4.7.
Lei Jun confirms Xiaomi's desktop AI agent MiClaw in development, MiMo-V2-Pro launches across all platforms
At the 2026 China Development High-level Forum, Xiaomi Group's Lei Jun confirmed that the long-awaited desktop version of the AI agent "MiClaw" (crab) is now on the development roadmap. Xiaomi had already launched a limited closed beta for the mobile
OpenAI Restarts Robot Business, Automan Seeks Engineers for Infrastructure R&D
On June 1st, OpenAI CEO Sam Altman announced on social media that the company is re-entering the robotics field, releasing job openings for the OpenAI Robotics team. The company is hiring full-stack hardware, operations, systems, and machine learning





Home






