The Agent Skills Directory

[SAFE]: The skill implements professional legal workflows for French, EU, and ECHR law. It is designed to work in environments like Cowork or with 'computer use' enabled to manage legal documents and verify citations.
[EXTERNAL_DOWNLOADS]: The skill includes Python scripts (doctrine_search.py, hal_search.py, verify_links.py) that perform network requests to retrieve legal data and academic references. These requests target well-known and official domains such as Légifrance (French government), EUR-Lex (European Union), HUDOC (ECHR), and HAL (French National Open Archive). According to the trusted source rules, these are legitimate operations for the skill's purpose.
[COMMAND_EXECUTION]: Several Python scripts are used to automate legal research and citation verification. These scripts use the Python standard library to perform HTTP requests, parse XML/Docx files, and manage JSON-based journals of legal references. No arbitrary or dangerous command execution (like piping remote scripts to a shell) was found.
[PROMPT_INJECTION]: The instructions in SKILL.md and principes-cardinaux.md contain rigorous 'anti-hallucination' rules. These rules are safety-positive, instructing the agent to never cite a law or case without first finding it in an official database and verifying that the content supports the assertion. No bypasses or safety-neutralizing injections were detected.
[DATA_EXFILTRATION]: While the skill reads user-provided documents (contracts, legal pieces), there is no evidence of exfiltration to untrusted domains. Network traffic is strictly mapped to the search APIs of the legal databases required for the task.
[INDIRECT_PROMPT_INJECTION]: The skill has an attack surface for indirect prompt injection as it reads untrusted user documents. However, this is mitigated by a mandatory pre-delivery checklist (Step 11) and a 'journal of references' system that requires the agent to verify every single citation against external official sources before delivery. This prevents the agent from being coerced into citing non-existent or malicious legal precedents.

assistant-juridique-fr