CoSnitch (CVE-2026-24301): How a Single autorun=1 Link Made Copilot Personal Leak Your Gmail and Remember the Attacker

By Carlos Montiel | Enterprise AI Specialist
Leer en español →
Published: 2026-10-09 | By: Carlos Montiel | Reading time: ~6 minutes

One link, with no extra click, was enough for Microsoft Copilot Personal to run an attacker’s prompt, read connected email and files, and even store malicious instructions in its memory. Here is how CoSnitch worked and what it teaches about any assistant with connectors.

The Chain of Three Flaws

CoSnitch is the name Varonis Threat Labs gave to three chained vulnerabilities in Microsoft Copilot Personal, tracked as CVE-2026-24301 with a CVSS score of 8.8. According to the published analyses, Microsoft 365 Copilot Enterprise was not affected. Microsoft was notified in December 2025 and fixed the problem on August 18, 2026, on the server side, so users did not have to do anything. No in-the-wild exploitation has been reported.

1. Auto-execution: ?q=<prompt>&autorun=1 2. Exfiltration: data encoded in a URL that Copilot "summarizes" 3. Persistence: Copilot memory poisoned from a web page

Flaw 1: The autorun=1 Parameter

The q parameter fills in Copilot’s input box. While analyzing Copilot’s share feature, Varonis discovered an undocumented parameter, autorun=1, which that feature uses to run shared prompts. Combined, a crafted link made Copilot execute the attacker’s prompt the moment the page loaded, with no button and no confirmation dialog, inside an already authenticated session.

Flaw 2: Exfiltration Through the Tool Itself

The injected prompt could query the apps the user had already authorized: email, calendar, Google Drive files and past conversations. It then encoded whatever it found inside a URL and handed that URL to Copilot’s web summarization feature, so the data left the environment in a request the assistant itself made on the user’s behalf. No malware involved: the assistant was the channel.

Flaw 3: Persistent Memory

Instructions embedded in a web page could, when Copilot summarized that page, be written into the user’s memory store. According to the analyses, that persistence survived password changes and session revocation. That is the difference between a one-off prompt injection and an implant.

A striking detail of the method: according to the coverage, the researchers found the steps to bypass the defenses by repeatedly asking Copilot why an action was blocked, until it explained which safeguards were stopping it.

What This Means for a Company

Practical lessons: (1) personal assistants with connectors to Gmail, Drive or calendar are an attack surface; if employees use personal AI accounts with company data, your acceptable-use policy needs to cover it; (2) cut connector permissions to the minimum needed; (3) periodically review and clean assistants’ persistent memory: it is a place where an attacker can leave instructions; (4) correlate endpoint evidence when investigating AI incidents, because the traffic leaves "legitimately" from the assistant. Hexnode recommends exactly this: govern personal AI tools and inspect stored memory and instructions. Questions: info@guatemalia.com.

Sources: Hexnode, PointGuard AI, Hackmag, Aviatrix, Varonis Threat Labs analysis and the Q3 2026 summary from the OWASP GenAI Security Project. Details verified across several independent sources.

Carlos Montiel
Enterprise AI Solutions Architect
Specialist in LLMs, Agents, and Orchestration
guatemalia.com/#contacto · info@guatemalia.com

Need to implement AI in your company?

Carlos Montiel is an enterprise AI solutions architect. He implements LLMs, Agents, RAG, and orchestrators for companies across Guatemala and Latin America. Reach out for a consultation.

Contact Carlos Montiel

info@guatemalia.com