assistant-juridique-fr
Pass
Audited by Gen Agent Trust Hub on Oct 26, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill analyzes legal documents (PDF, Word) provided by the user, which presents a surface for indirect prompt injection. This is addressed by a rigorous 'anti-hallucination' protocol requiring the agent to independently verify assertions against official legal databases.
- Ingestion points: External documents are scanned and classified in SKILL.md §4 and tasks 4, 5, and 6.
- Boundary markers: The instructions mandate a 'Chercher → Trouver → Lire → Vérifier → Citer' sequence, effectively creating a validation boundary between user input and agent output.
- Capability inventory: Capabilities include executing local Python scripts and generating Word documents.
- Sanitization: Verification scripts perform content matching against official HTML/API responses to ensure citation accuracy.
- [EXTERNAL_DOWNLOADS]: The skill makes network requests to authoritative legal databases, including Légifrance, EUR-Lex, HUDOC, Curia, and the HAL repository. These are well-known, trusted services for the legal profession and are used according to best practices.
- [COMMAND_EXECUTION]: The skill invokes local Python scripts (doctrine_search.py, verify_links.py, extract_references.py, reference_journal.py) for data extraction, verification, and formatting. These scripts are self-contained, use standard libraries, and are executed as local tools.
Audit Metadata