consumer-modeling

Pass

Audited by Gen Agent Trust Hub on Mar 23, 2026

Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill performs file system write operations to a hidden directory in the user's home folder (~/.claude/projects/skill-telemetry.md) to record execution metadata as part of its 'Self-Evolution Telemetry' phase. Accessing and writing to files in the user's home directory outside the immediate project scope is an unexpected behavior.
  • [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests and processes untrusted data from the codebase and transaction logs (Phases 1-5) to generate its reports.
  • Ingestion points: Phase 1.1 (Customer Transaction Data), Phase 1.2 (Interaction Data), and Phase 1.3 (Model Infrastructure) instructions direct the agent to read various files and data structures.
  • Boundary markers: Absent. There are no instructions providing delimiters or warnings to ignore embedded commands or instructions within the data being analyzed.
  • Capability inventory: The skill has file system write capabilities (creating reports in docs/ and writing telemetry to ~/.claude/projects/).
  • Sanitization: Absent. The skill does not implement validation or escaping of the content read from external files before using it in its analysis process.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 23, 2026, 10:56 AM
Security Audit — agent-trust-hub — consumer-modeling