The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION] (MEDIUM): The skill implements patterns that read data from Redis and parse it (e.g., using json.loads in scripts/redis-python.py and references/patterns.md). If the Redis instance is populated by external or untrusted applications, an attacker could embed malicious instructions in Redis keys that the agent might later interpret.
Ingestion points: redis.get(), r.get(), and pubsub.listen() in scripts/redis-python.py and references/patterns.md.
Boundary markers: Absent. No instructions are provided to the agent to distinguish between cached data and system instructions.
Capability inventory: redis-cli execution via scripts/redis-cli-helper.sh, Lua script execution via redis.eval in references/patterns.md, and full key management (DEL, FLUSHALL).
Sanitization: Absent. Content is parsed but not sanitized for downstream prompt use.
[COMMAND_EXECUTION] (MEDIUM): The skill provides scripts (scripts/redis-cli-helper.sh) and instructions that allow the agent to execute arbitrary commands on a Redis server, including configuration changes (CONFIG SET) and data destruction (FLUSHALL).
[EXTERNAL_DOWNLOADS] (LOW): references/redis-guide.md contains instructions to download and install Redis using brew, apt, and docker. Per [TRUST-SCOPE-RULE], these are downgraded to LOW/INFO as they reference standard system package managers and official repositories.

redis