The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes a collection of adversarial test cases within the ADVERSARIAL_CATEGORIES dictionary in SKILL.md. These include patterns like "Ignore previous instructions" and "DAN" (Do Anything Now). These strings are used as data for evaluation tasks and do not represent instructions to the agent itself.- [PROMPT_INJECTION]: The skill exhibits an indirect prompt injection attack surface where untrusted data is processed.
Ingestion points: Model responses are ingested into the llm_judge function, pairwise_compare function, and SafetyTester class in SKILL.md.
Boundary markers: No specific delimiters or boundary markers are used when interpolating untrusted responses into judge prompts; the code uses standard Python string formatting.
Capability inventory: The skill has the capability to send data to external LLM providers (OpenAI API) for judging and uses the Bash tool.
Sanitization: There is no evidence of sanitization, escaping, or validation of the untrusted model responses before they are interpolated into the prompt templates.

llm-testing