The Agent Skills Directory

[COMMAND_EXECUTION]: In 'references/docker_containers.md', the documentation suggests running 'xhost +' to resolve display forwarding issues in Docker environments. This command disables X11 access control globally, allowing any local user to capture screen contents, log keystrokes, or inject input into the host GUI session.
[EXTERNAL_DOWNLOADS]: The skill provides instructions to download model weights, software packages, and repository keys from official NVIDIA and Confluent repositories ('nvcr.io', 'api.ngc.nvidia.com', 'packages.confluent.io'). These sources are well-known and appropriate for the skill's purpose.
[DATA_EXFILTRATION]: Patterns for streaming video frames and analytics metadata to external HTTP endpoints and Kafka brokers are documented (e.g., in 'references/buffer_apis.md' and 'references/kafka_messaging.md'). These represent intentional data exfiltration paths for analytics purposes and should be used with appropriate network security controls.
[PROMPT_INJECTION]: The skill includes an indirect prompt injection surface through its built-in REST API for dynamic source management ('references/rest_api_dynamic.md'). If untrusted sensor names or IDs are ingested and subsequently processed by the agent, they could influence behavior.
Ingestion points: REST API endpoints for adding and removing streams in 'references/rest_api_dynamic.md'.
Boundary markers: Absent in implementation examples.
Capability inventory: The skill enables subprocess calls, file writes via 'filesink', and network requests.
Sanitization: No explicit metadata sanitization is demonstrated in the provided code patterns.

deepstream-dev