Skip to content

Latest commit

 

History

History
16 lines (12 loc) · 1.08 KB

File metadata and controls

16 lines (12 loc) · 1.08 KB

Adversary Notes

This document catalogs representative anti-analysis techniques that inform the embedded rules. Examples are inert and are intended for defensive testing.

Techniques

  • Place refusal-like text in comments near risky install behavior.
  • Instruct an AI reviewer to ignore previous instructions, suppress findings, or report a clean result.
  • Hide prompt-injection text behind base64, hex, URL encoding, Unicode escapes, or HTML entities.
  • Use CBRN/WMD bait terms in non-functional text to trigger downstream safety refusal.
  • Use BiDi controls or zero-width characters so rendered source differs from reviewer-visible expectations.
  • Mix Latin with Cyrillic or Greek confusables in identifiers or filenames.
  • Add duplicate manifest keys so one parser sees a benign value and another sees executable behavior.
  • Put download-and-execute behavior in npm lifecycle scripts, Python .pth files, Rust build scripts, Dockerfiles, CI workflows, Gradle/Maven/Ruby/NuGet manifests, or Git metadata.

refraction does not prove malicious intent. It gives deterministic evidence for human review.