session-anonymizer
Pass
Audited by Gen Agent Trust Hub on Sep 19, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: Fetches the
privacy-filterpackage from OpenAI's official GitHub repository to provide a specialized anonymization layer.\n- [COMMAND_EXECUTION]: Executes theopfCLI tool and the systemopensslutility via subprocess for PII detection and optional output encryption. These calls use static argument lists which prevent shell injection.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted transcript data which is passed directly into a local LLM prompt to identify PII, creating a potential surface for prompt injection within the local processing stream.\n - Ingestion points: Untrusted text is read from files or stdin in
scripts/anonymize.py.\n - Boundary markers: Absent; the input text is appended directly to the extraction prompt instructions.\n
- Capability inventory: The skill has capabilities to read/write files, execute specific shell commands (opf, openssl), and communicate with the local network (Ollama API).\n
- Sanitization: Input text is not sanitized or escaped before being included in the prompt.
Audit Metadata