The Agent Skills Directory

[SAFE]: The skill consists solely of markdown instructions and a notice file. It does not include any scripts (Python, JavaScript), binaries, or other executable components.\n- [DATA_EXFILTRATION]: No network operations or data exfiltration patterns were identified. The skill configuration explicitly defines data-residency: local and pii-egress: none, and the instructions proactively advise users to pseudonymize sensitive documents before processing.\n- [PROMPT_INJECTION]: The skill defines a structured workflow using multiple roles (Builder, Attacker, Synthesizer, Verifier). These instructions are aimed at guiding the agent's reasoning process and do not contain patterns designed to bypass safety filters, extract system prompts, or override core instructions.\n- [SAFE]: The skill processes external document data, creating a potential surface for indirect prompt injection. However, the workflow mitigates this risk through isolation tiers (ISOLATED vs. AUGMENTED), which restrict a reviewer's access to original sources to prevent bias or manipulation, and a mandatory 10-point verification check that includes mechanical citation grounding.

adversarial-legal-review-en