The Agent Skills Directory

[DATA_EXFILTRATION]: The skill manages files containing Personally Identifiable Information (PII), such as research/audience/ledger.csv and creators/roster.csv. These files store emails, names, and profile URLs for marketing tracking.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the user when they report outcomes. * Ingestion points: User-provided text, pasted content, or uploaded files containing outcome reports (e.g., "Paste, upload, or tell me in a sentence"). * Boundary markers: Absent for the user-provided data. * Capability inventory: File read/write operations, shell/code execution for hashing, and interaction with memory skills (save-learning, save-to-brain). * Sanitization: Instructions mandate anonymizing PII during suppression, but there are no measures to sanitize or escape user input to prevent instructions embedded in reports from influencing the agent.
[COMMAND_EXECUTION]: The skill requires the use of shell or code execution tools to compute SHA-256 hashes for data anonymization. * Evidence: Instructions state "Compute the hash with the host's code tool (shell or code execution)" to generate the key_hash field for suppression.

outcomes-ledger