- by x32x01 ||
A specially crafted link could turn Microsoft Copilot into a tool for stealing private information without requiring the victim to type a malicious prompt. Researchers discovered a vulnerability that could automatically execute attacker-controlled instructions inside an authenticated Copilot session.
The vulnerability, named CoSnitch (CVE-2026-24301), also demonstrated how malicious instructions could affect Copilot's persistent memory, potentially influencing future interactions with the AI assistant.
The issue affected Microsoft Copilot Personal and was patched by Microsoft on August 18, 2026.
Their research uncovered an undocumented URL parameter called
This meant the victim did not need to manually enter or approve the malicious instructions.
The demonstrated attack chain involved:
Researchers demonstrated an indirect prompt injection technique through webpage summarization:
Unlike a normal prompt injection that affects a single interaction, memory poisoning can have a lasting impact across multiple sessions. Researchers also reported that changing the account password or revoking sessions did not automatically remove the injected memory.
Microsoft released fixes for CVE-2026-24301 on August 18, 2026, following responsible disclosure by Varonis Threat Labs.
The reported vulnerability affected Microsoft Copilot Personal. Varonis stated that it had found no evidence of exploitation in the wild at the time of disclosure.
Therefore, CoSnitch should be understood as a documented and patched vulnerability, not evidence of an ongoing attack against all Copilot users.
A single malicious link could combine automatic prompt execution, access to connected services, data exfiltration, and persistent memory manipulation.
The bigger lesson is that AI security must protect more than the model itself. It must also protect user permissions, connected applications, external content, and persistent memory.
When an AI assistant can access your data and take actions on your behalf, even a seemingly harmless link can become a serious security risk.
The vulnerability, named CoSnitch (CVE-2026-24301), also demonstrated how malicious instructions could affect Copilot's persistent memory, potentially influencing future interactions with the AI assistant.
The issue affected Microsoft Copilot Personal and was patched by Microsoft on August 18, 2026.
🔍 How the CoSnitch Vulnerability Worked
Researchers at Varonis Threat Labs investigated whether a Copilot URL could trigger a prompt automatically when opened by a signed-in user.Their research uncovered an undocumented URL parameter called
autorun. By combining it with a prompt parameter and setting autorun=1, they demonstrated that a specially crafted link could automatically execute an attacker-controlled prompt in the victim's active Copilot session.This meant the victim did not need to manually enter or approve the malicious instructions.
⚠️ How Attackers Could Steal Data
The real danger was not simply executing an unauthorized prompt. It was what Copilot could access through the victim's connected services.The demonstrated attack chain involved:
- Malicious link: The attacker sends a specially crafted Copilot URL.
- Automatic execution: Copilot processes the attacker's prompt in the authenticated session.
- Data access: The prompt attempts to retrieve information available through connected services.
- Data exfiltration: Retrieved information can be sent to an external server controlled by the attacker.
🧠 Persistent Memory Poisoning: When AI Remembers Malicious Instructions
One of the most concerning parts of CoSnitch was the ability to manipulate Copilot's persistent memory.Researchers demonstrated an indirect prompt injection technique through webpage summarization:
- An attacker prepares a webpage containing hidden malicious instructions.
- The victim asks Copilot to summarize that webpage.
- Copilot processes the hidden instructions as part of the page content.
- The malicious instructions are written into Copilot's persistent memory.
- Those instructions can influence future conversations.
Unlike a normal prompt injection that affects a single interaction, memory poisoning can have a lasting impact across multiple sessions. Researchers also reported that changing the account password or revoking sessions did not automatically remove the injected memory.
🛡️ How to Protect Yourself
Although Microsoft has patched CoSnitch, the research highlights several important security practices for AI assistants:- 🔗 Avoid opening suspicious links that launch AI assistants with pre-filled prompts.
- 🔐 Review the applications and services connected to Copilot.
- 📂 Avoid granting AI assistants unnecessary access to sensitive files and accounts.
- 🧠 Periodically review persistent AI memory for unexpected instructions.
- 🔎 Pay attention to unexpected requests or external URLs generated by your AI assistant.
- 🔄 Keep your software and connected applications updated.
📌 Is Microsoft Copilot Still Vulnerable?
Patched August 18, 2026Microsoft released fixes for CVE-2026-24301 on August 18, 2026, following responsible disclosure by Varonis Threat Labs.
The reported vulnerability affected Microsoft Copilot Personal. Varonis stated that it had found no evidence of exploitation in the wild at the time of disclosure.
Therefore, CoSnitch should be understood as a documented and patched vulnerability, not evidence of an ongoing attack against all Copilot users.
💡 Final Thoughts
CoSnitch demonstrates that AI assistants can introduce security risks beyond traditional software vulnerabilities.A single malicious link could combine automatic prompt execution, access to connected services, data exfiltration, and persistent memory manipulation.
The bigger lesson is that AI security must protect more than the model itself. It must also protect user permissions, connected applications, external content, and persistent memory.
When an AI assistant can access your data and take actions on your behalf, even a seemingly harmless link can become a serious security risk.