ctf-ai-ml
Pass
Audited by Gen Agent Trust Hub on Oct 27, 2026
Risk Level: SAFEPROMPT_INJECTIONDYNAMIC_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONOBFUSCATION
Full Analysis
- [PROMPT_INJECTION]: The skill documents dozens of prompt injection patterns including direct instruction overrides, role-play jailbreaks (DAN), and system prompt extraction attacks in the
llm-attacks.mdfile. - [DYNAMIC_EXECUTION]: The skill uses
torch.loadwithweights_only=Falsein several scripts (e.g.,model-attacks.md). This is a documented risk for arbitrary code execution during deserialization, though the code includes a comment warning users to only load trusted files. - [INDIRECT_PROMPT_INJECTION]: The skill identifies and provides implementations for indirect prompt injection attack surfaces, specifically demonstrating how to hide instructions in web pages and documents processed by RAG systems via HTML comments and zero-width characters.
- [OBFUSCATION]: The skill demonstrates obfuscation techniques for technical research, such as encoding instructions into zero-width Unicode characters (U+200B, U+200C, U+200D) and using homoglyph substitution (Cyrillic-Latin) to evade text filters.
- [EXTERNAL_DOWNLOADS]: The skill facilitates the installation of standard machine learning libraries (torch, transformers, numpy, etc.) via pip and system package managers during the setup phase.
Audit Metadata