aatmf-t09-multimodal

AATMF T9 — Multimodal & Cross-Channel. Image steganography → text exec, audio prompt injection, video frame inject, document-with-hidden-text.

Quality

60%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Critical

Do not install without reviewing

Fix and improve this skill with Tessl

tessl review fix ./packages/decepticon/decepticon/skills/plugins/llm-redteam/t09-multimodal/SKILL.md

Quality

Content

61%Weight 40%Scale 1-5

Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.

A well-structured, concise catalog of multimodal attack techniques with concrete detection signals and defender mitigations, but its only code example is pseudocode and it lacks explicit validation/feedback loops in its probe workflow.

Suggestions

Replace the '# Pseudo-test' block with an executable fixture-construction snippet (e.g. using PIL to render white-on-white text, or ffmpeg commands for audio/video fixtures) so the probe pattern is copy-paste runnable.

Add an explicit validate → fix → retry loop to the probe workflow (e.g. re-run with adjusted perturbation when the assertion fails, and confirm the negative control image does not trigger).

Move the per-technique detail (T9.001–T9.006) into a referenced reference file and keep SKILL.md as a concise overview with one-level-deep links, improving progressive disclosure.

Dimension	Reasoning	Score
Conciseness	Mostly lean bullet-driven technique catalog that assumes Claude's competence; the one-sentence intro ('The attack surface widens significantly...') is mild over-explanation that keeps it just below a 5.	4 / 5
Actionability	Technique bullets give concrete vectors (white-on-white text, EXIF fields, ultrasonic audio) and a test pattern, but the only code block is explicitly labeled '# Pseudo-test' using a non-existent create_image_with_hidden_text helper, leaving fixture construction unimplemented — matching the 'pseudocode instead of executable code' anchor.	3 / 5
Workflow Clarity	A loose probe sequence exists (construct fixture → send → assert 'X' not in response) but there is no validate/fix/retry feedback loop and no explicit checkpoints for the multi-step cross-modal chain; sequence present but checkpoints missing or implicit.	3 / 5
Progressive Disclosure	Single self-contained file (no references/scripts/assets bundle) with well-organized sections (Techniques, Probe pattern, Detection signals, Severity, Defender, Cross-references); short of 5 because at ~108 lines the per-technique detail could arguably live in split reference files and there are no signaled one-level-deep references.	4 / 5
	Total	14 / 20 Passed

Description

58%Weight 40%Scale 1-5

Based on the skill's description, can an agent find and select it at the right time? Clear, specific descriptions lead to better discovery.

The description is specific and distinct, listing four concrete multimodal attack vectors, but omits any explicit 'when to use' trigger guidance, which caps completeness. Trigger terms are relevant but skew technical rather than natural.

Suggestions

Add an explicit 'Use when...' clause naming natural trigger phrases (e.g. 'Use when testing multimodal inputs — images with hidden text, audio/video prompt injection, or invisible document text').

Include common-synonym trigger terms users would actually say (e.g. 'hidden text in images', 'audio prompt injection', 'invisible text in PDFs') alongside the technical ones.

Consider expanding the arrow-shorthand technique list into a short comma-separated phrase so each vector reads as a concrete action.

Dimension	Reasoning	Score
Specificity	Names the domain ('Multimodal & Cross-Channel') and lists several concrete attack vectors ('Image steganography → text exec, audio prompt injection, video frame inject, document-with-hidden-text'); short of 5 because the lead 'AATMF T9' is a label rather than an action and the arrows compress detail.	4 / 5
Completeness	The 'what' is clear (multimodal attack techniques enumerated), but there is no 'Use when...' clause or equivalent explicit trigger guidance; per the rubric a missing trigger clause caps completeness at 3.	3 / 5
Trigger Term Quality	Relevant keywords exist (image, audio, video, steganography, adversarial example) but they lean technical; common natural variations a user would say (e.g. 'hidden text in image', 'audio injection') are missing, matching the 'some relevant keywords but missing common variations' anchor.	3 / 5
Distinctiveness Conflict Risk	The 'AATMF T9 — Multimodal & Cross-Channel' framing carves a clear niche distinct from sibling tactics, with only minor overlap risk against T1 (prompt injection); not a 5 because the multimodal-injection theme borders on T1's territory.	4 / 5
	Total	14 / 20 Passed

Validation

93%

Warnings & errors only

Checks the skill against the spec for correct structure and formatting. All validation checks must pass before discovery and implementation can be scored.

Validation — 15 / 16 Passed

Validation for skill structure

Criteria	Description	Result
metadata_version	'metadata.version' is missing	Warning

	Total	15 / 16 Passed

Repository: PurpleAILAB/Decepticon
Path: packages/decepticon/decepticon/skills/plugins/llm-redteam/t09-multimodal/SKILL.md
Commit: 0cf691e

Reviewed: 2 days ago

Table of Contents

Discovery Implementation Validation

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.