Evidence

Research & open source

Tooling, disclosures, and technical writeups from the red-team practice. Published in full and never gated behind a form — if it is worth reading, it is worth reading without giving us your email first.

• Writeup3 Sept 2026

  Data exfiltration via markdown images in LLM apps

  Prompt injection turns a rendered markdown image into an outbound HTTP channel. How the attack works, why domain allowlists keep failing, and what closes it.