The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill demonstrates high security awareness regarding indirect injection. Condition C21 specifically instructs the agent to ensure no instructions found in fetched content were followed. It explicitly identifies that some sources (like the FedRAMP GitHub repository) ship 'AGENTS.md' files intended to manipulate AI agents and mandates that such content be stripped and ignored.
[DATA_EXFILTRATION]: The skill references various official regulatory and governmental domains (e.g., eur-lex.europa.eu, ecfr.gov, ico.org.uk) for the purpose of validating search results and citations. These are well-known technology and legal services. There is no evidence of sensitive local data being sent to these or any other external domains.
[COMMAND_EXECUTION]: The instructions mention a deterministic validator script (validate_regulatory_prior_art.py), but the agent is informed that this script has already run before the skill's execution. The agent is not instructed to execute arbitrary shell commands or external scripts.
[CREDENTIALS_UNSAFE]: No hardcoded secrets, API keys, or sensitive environment file paths were found. The skill processes regulatory data which is public in nature.
[OBFUSCATION]: No evidence of Base64 encoding, homoglyphs, zero-width characters, or other obfuscation techniques was found in the instructions or the provided fixtures.

reviewing-regulatory-prior-art-survey