The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accesses the local directory ~/.claude/projects/ to read and write execution metadata in the skill-telemetry.md file. While this is used for self-evolution and project memory, accessing hidden configuration directories is a sensitive operation that can expose environment-specific metadata.
[DATA_EXFILTRATION]: The skill is designed to search for and read highly sensitive pharmaceutical data assets, including historian databases (OSIsoft PI, InfluxDB), MES/EBR systems (Syncade, PAS-X), and LIMS data. This access is inherent to the skill's primary function of yield prediction but involves high-value data.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it processes untrusted data from the filesystem and incorporates it into analytical reports.
Ingestion points: Data enters the context during Phase 1 (Step 1.1) from external historians, batch records, LIMS, and PAT sensors.
Boundary markers: No delimiters or warnings are used to separate untrusted data from the agent's instructions.
Capability inventory: The skill has file-writing capabilities to docs/yield-prediction-analysis.md and skill-telemetry.md (via shell appending).
Sanitization: There is no evidence of sanitization, escaping, or validation of the ingested process data before it is processed or written to output files.

yield-prediction