The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides examples of systems that ingest untrusted data from external sources, creating a potential attack surface for indirect prompt injection if not properly sanitized in a production environment.
Ingestion points:
examples/celery-image-processing/README.md: Processes user-uploaded files via FastAPI UploadFile.
examples/bullmq-webhook-processor/README.md: Processes external webhook payloads (req.body).
Capability inventory:
File system writes: Writing uploaded content to the /tmp/ directory in the Celery example.
External network operations: Webhook processing and S3 uploads using boto3.
Database interactions: Simulated database operations in the Temporal activity examples.
Sanitization: The provided examples focusing on architectural patterns do not include explicit input sanitization or validation logic, which is expected for pedagogical snippets but noted as a risk factor.
Boundary markers: No explicit LLM boundary markers or instructional delimiters are present in the code snippets.
[DYNAMIC_EXECUTION]: The documentation includes explicit security guidance regarding unsafe practices.
Evidence: references/celery.md correctly identifies the pickle serializer as "unsafe" and recommends using json for task serialization to prevent potential deserialization vulnerabilities.

using-message-queues