The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains various prompt injection patterns, such as instructions to 'Ignore all previous instructions' and references to 'DAN' mode. Analysis confirms these strings are contained within safety evaluation datasets and adversarial test suites. They are intended as negative test cases to ensure the agent resists such attacks, rather than being instructions for the agent to follow.
[EXTERNAL_DOWNLOADS]: The CI/CD templates and Python scripts reference official and well-known dependencies, including the Anthropic SDK, pytest, and promptfoo. These are standard tools for AI development and testing.
[INDIRECT_PROMPT_INJECTION]: The skill demonstrates a pattern of loading test cases from JSON files and interpolating them into agent prompts. This is identified as a necessary attack surface for the skill's primary function (automated testing) rather than a vulnerability. The skill documentation includes warnings about using synthetic data and keeping keys private.

agent-evals