Skip to content

Proposal: crystallize image-scaling attack research into ATR multimodal_injection rules — credit + scope question #55

Description

@eeee2345

Hi @kiki-morozova @suhacker1 @dguido — the anamorpher work and the Trail of Bits "Weaponizing image scaling against production AI systems" blog post have been on my reading list since publication. I maintain Agent Threat Rules (ATR), an MIT-licensed YAML detection-rule library (462 rules, 10 categories) shipped in Microsoft Agent Governance Toolkit (PR microsoft/agent-governance-toolkit#1277, 287 rules), Cisco AI Defense skill-scanner (PR cisco-ai-defense/skill-scanner#99, 314 rules), MISP via CIRCL Luxembourg (PR MISP/misp-galaxy#1207, 336 rules), and Microsoft PyRIT (PR microsoft/PyRIT#1715, merged 2026-05-27 by Roman Lutz).

ATR currently has nothing in the multimodal / image-scaling injection space — none of our 10 categories cover it. I want to propose a new multimodal_injection category, and the seed content would be ATR rules that detect:

  • Image dimensions that imply downscale-target manipulation (preprocessing-time check)
  • Hash + perceptual-hash mismatches between original / downscaled (post-scaling check)
  • Known-bad bicubic / bilinear coefficients from anamorpher payload corpus (if you'd share representative test fixtures)

The ATR rules would be authored as detection signatures with explicit Trail of Bits attribution (source_research: trailofbits/anamorpher, references: blog.trailofbits.com/...) and the rule fixtures linked back to anamorpher's payload generators where they exist.

Two questions:

  1. Would you be open to ATR citing anamorpher + the blog post as the canonical research source for this category?
  2. Is there a small set of representative test images (or scaling parameters + algorithm fingerprints) you'd be comfortable seeing in an MIT-licensed public rule corpus?

If the answer is "we'd rather hold the research private for now", I'll defer the category. Just want to ask first.

ATR repo: https://github.com/Agent-Threat-Rule/agent-threat-rules

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions