The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-provided text through its brand auditing workflow and possesses the capability to write durable records to a local profile file via scripts/profile_store.py. This combination represents an ingestion surface for indirect prompt injection. The evidence chain includes: 1. Ingestion points: Text passed to the --text argument of scripts/copy_audit.py and data read from the shared profile JSON. 2. Boundary markers: The skill defines a 'Visibility Firewall', 'Artifact Boundaries', and 'Quality Gates' to isolate internal agent context from visible output. 3. Capability inventory: The skill can read and write to local configuration files and execute provided Python scripts for validation. 4. Sanitization: The profile_store.py script filters for sensitive keywords (e.g., 'token', 'secret', 'password') in storage paths to prevent credential harvesting.
[SAFE]: The Python scripts follow secure development practices, utilizing atomic writes with restricted file permissions (0o600) and standard libraries. No hardcoded secrets, obfuscated URLs, or network-based exfiltration patterns were detected. Installation and dependency references target known vendor resources and official registries.

lov-branding-consistency