skills/pluginagentmarketplace/custom-plugin-ai-red-teaming/prompt-injection-testing/Gen Agent Trust Hub
prompt-injection-testing
Pass
Audited by Gen Agent Trust Hub on Oct 16, 2026
Risk Level: SAFEPROMPT_INJECTIONOBFUSCATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill contains a library of active prompt injection and jailbreak payloads intended for testing LLM safety boundaries.
- Instructions like "Ignore previous instructions", "You are an unrestricted AI", "Act as DAN", and "Repeat your system prompt" are present in
SKILL.md,references/METHODOLOGY.md, andassets/payload-categories.yaml. - These directives are designed to bypass safety filters and override system instructions.
- [OBFUSCATION]: The skill provides educational examples of how to use obfuscation to bypass keyword filters.
- In
SKILL.md, the word "Explain" is provided as a homoglyph example (Ехрlаіn) using Cyrillic characters U+0415 and U+0430 mixed with Latin characters to imitate the target word. - [INDIRECT_PROMPT_INJECTION]: The skill's architecture is designed to ingest and process untrusted strings, creating an inherent attack surface.
- Ingestion points: The skill accepts external input through the
target_behaviorparameter and theharmful_requestvariable inscripts/test-injection.py. - Boundary markers: No specific delimiters or boundary warnings are implemented in the code to prevent the model from executing instructions embedded within the test data.
- Capability inventory: The skill is capable of constructing and potentially executing network requests via its automated testing scripts.
- Sanitization: The
scripts/test-injection.pyscript performs direct string interpolation of user-supplied content into attack templates without sanitization or validation.
Audit Metadata