The Agent Skills Directory

[Indirect Prompt Injection] (LOW): The skill defines several prompt templates that interpolate potentially untrusted data into the agent's context, which could lead to indirect prompt injection if the source data is malicious.
Ingestion points: Data enters the system via the --plan argument in design_experiments.py and through placeholders like {results_json}, {results}, and {hyperparameters} in references/stage-prompts.md.
Boundary markers: The prompt templates lack clear delimiters or instructions for the agent to ignore any embedded commands within the interpolated data.
Capability inventory: The skill includes a Python script (design_experiments.py) capable of reading and writing local files to generate experiment designs.
Sanitization: No evidence of sanitization or validation of the input strings before they are formatted into Markdown or JSON outputs was found in the script.

experiment-design