Source: Markdown exfiltration (Simon Willison tag archive)

Willison's running catalogue of products leaking data through injected Markdown images or links, and why CSP and URL allowlists keep failing.

Contents

Summary

The canonical prompt-injection exfiltration channel: injected text tells the model to emit a Markdown image whose URL contains private data; when the client renders it, the browser sends the data to the attacker’s server. Link-based variants need one click.

Key claims

  • Affected products listed include Mistral LeChat (2024-10), the Claude iOS app (2024-12), GitLab Duo (2025-05), Microsoft 365 Copilot “EchoLeak” (2025-06), Salesforce Agentforce (2025-09), Superhuman AI (2026-01) and Microsoft Copilot Cowork (2026-05); Slack AI (2024-08) used links.
  • Fixes and how they failed: CSP img-src allowlists defeated by an allowed domain that stores attacker data (Google Forms on docs.google.com), an expired allowlisted domain, or an open redirect; external-image filters missed reference-style Markdown (EchoLeak); allowlists that included webhook.site or dozens of package-manager domains.
  • Restricting outbound HTTP to GET/HEAD/OPTIONS and disabling link unfurling help.
  • In application security, 99% is a failing grade; remove the exfiltration leg instead.

Feeds

Prompt injection · LLM security risks

Suggest changes →