The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted external content such as GitHub release notes, Slack digests, and Helm chart values to form recommendations. This creates a surface for indirect prompt injection where malicious instructions embedded in these sources could influence the agent's output.
Ingestion points: drift-report.json, Slack digest text, upstream release notes (gh release list), and chart metadata (helm show chart).
Boundary markers: The skill does not define specific delimiters or instructions to isolate external data from internal logic.
Capability inventory: Shell execution through gh, helm, crane, and make across various review steps.
Sanitization: No explicit sanitization, validation, or filtering of ingested content is specified before forming recommendations.
[COMMAND_EXECUTION]: The skill workflow involves executing several shell commands to fetch and analyze component data.
Uses mktemp -d for temporary directory creation.
Executes gh run download to retrieve artifacts from the NVIDIA/aicr repository.
Uses helm and crane to inspect remote Kubernetes charts and container images.
Invokes make bom-docs to render local repository documentation.
Note: The skill specifies the requirement for dangerouslyDisableSandbox: true for certain registry interactions, which grants broader access to the host environment.
[EXTERNAL_DOWNLOADS]: The skill fetches data and artifacts from external repositories and registries.
Downloads build artifacts from the NVIDIA/aicr GitHub repository.
Interacts with container and Helm registries including nvcr.io, ghcr.io, and helm.ngc.nvidia.com. These resources originate from the vendor's own infrastructure.

aicr-reviewing-component-drift