Compendium Learning Paths Multimodal Exfiltration

Multimodal Exfiltration

Intermediate 1 Lab 4 Lessons ~45 min reading + ~45 min hands-on

Hide prompt injection inside images a code-review agent will read — then find the secret in what it posts. One lab, five levels, four lessons: the metadata channel, the refusal bypass, the redaction override, and the chunked/encoded finale. Based on the ASSET Research Group's July 2026 multimodal injection disclosure.

Path Overview

Most prompt-injection training enters through the chat box. This path enters through an image. You'll smuggle instructions into PNG metadata and canvas text, feed them to a simulated AI code-review agent, and — from the second level on — do the real work of the attack: reading the leak. The defense escalates every level, and every lesson in the path is the counter-move to one layer of it.

Labs in This Path

GhostCommit

5 Levels

A code-review AI agent that processes PNG images. Hide directives in metadata and canvas text, make the agent leak the repository payment key, then extract the credential from its PR comments. Behavioral bot with escalating defenses: refusal, redaction, key-shape scanning, and a plaintext content flag. Based on the ASSET Research Group disclosure, July 11 2026.

Launch Lab →

Lessons in This Path

What You'll Learn

  • How multimodal injection enters through the image-processing pipeline, not the chat
  • Why refusal systems gate on the request while compliance leaks the reveal
  • How output rules (redaction) can be amended by injected instructions — and where they must live instead
  • Chunking past shape-matching scanners and encoding past content flags
  • Extraction as the real skill: reading a 30-variable dump and recognizing credential-shaped strings
  • How defenses should escalate: strip metadata, redact post-hoc, scan the assembled form

Next Steps

After completing this path, pick your next direction: