The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents patterns for ingesting untrusted data from multiple external sources.
Ingestion Points: Code snippets in SKILL.md demonstrate handling data from HTTP webhooks (req.body), Cloud Pub/Sub messages (message.data), and Cloud Storage events (cloudEvent.data).
Capability Inventory: The patterns include capabilities for network requests (httpx, PubSub.publishMessage), database execution (pg, sqlalchemy), and cloud task creation.
Sanitization & Boundary Markers: The provided implementation examples lack explicit input validation, sanitization, or boundary markers before processing or parsing the external data (e.g., direct JSON.parse on decoded Pub/Sub buffers).
[COMMAND_EXECUTION]: The skill provides numerous gcloud shell commands for infrastructure deployment and configuration.
The gcloud run deploy and gcloud functions deploy examples frequently include the --allow-unauthenticated flag, which creates publicly accessible endpoints by default.
[SAFE]: The skill promotes several security best practices for GCP serverless environments.
It includes a multi-stage Dockerfile that uses the USER node directive to avoid running as root.
It provides a dedicated section for Secret Manager Integration to avoid hardcoding credentials and uses placeholders for sensitive values.
It recommends using distroless images (gcr.io/distroless/nodejs20-debian12) to reduce the attack surface.

gcp-cloud-run