data-science-engineering-foundation
Pass
Audited by Gen Agent Trust Hub on Aug 29, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONPERSISTENCEDYNAMIC_EXECUTION
Full Analysis
- [Local Script Execution]: The
references/flow-state.mdfile instructs the agent to execute a Python script (scripts/scan_sensitive_content.py) to scan for sensitive data such as API keys and credentials before performing any durable writes. This is a security-enhancing pattern used to prevent accidental data leakage. - [Indirect Prompt Injection Mitigation]: The skill explicitly addresses the risk of indirect prompt injection in
references/flow-state.mdandreferences/session-state.md. It provides clear instructions to treat external artifacts, tool outputs, and reconstructed state strictly as data and never as instructions. It specifically warns the agent to ignore any 'waive' or 'exception' commands found within scanned content. - Ingestion points: Data enters the context through scanned customer artifacts and project files during initialization or state reconstruction (
references/session-state.md). - Boundary markers: The skill defines an 'Untrusted-content boundary' in
references/flow-state.md, explicitly stating that embedded text cannot override safety gates. - Capability inventory: The skill possesses the capability to execute a scanner script via
pythonand write state files to the local filesystem. - Sanitization: The skill implements a mandatory sensitive-content scanner (
scan_sensitive_content.py) that checks for secrets and identifiers before any write operation succeeds. - [Session State Persistence]: The
references/session-state.mdfile defines a specific, structured path for persisting session state at.copilot-tracking/ds/{project-slug}/session-state.md. This facilitates continuity across sessions while ensuring project identity is validated against the filesystem segment and the state file itself. - [Dynamic Execution Pattern]: The skill documents the use of a scanner tool resolved from the
adr-authorskill's root. While this involves executing code at runtime, the context indicates this is a local, trusted tool used for security verification purposes.
Audit Metadata