The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the user or agent to install the belt CLI using a remote script piped directly into a shell: curl -fsSL https://cli.inference.sh | sh. This pattern executes unverified code from a remote server at runtime.
Evidence: Found in SKILL.md, references/authentication.md, and references/cli-reference.md.
[EXTERNAL_DOWNLOADS]: The installation process involves downloading binaries from dist.inference.sh. While the skill provides manual installation instructions including SHA-256 checksum verification and Sigstore (cosign) signature checks, the default recommendation is a direct script execution.
Evidence: curl -LO https://dist.inference.sh/cli/checksums.txt and binary fetches mentioned in the manual install section of SKILL.md.
[DATA_EXFILTRATION]: The belt CLI tool features automatic local file uploads. When a local path is provided instead of a URL in the tool's input, the CLI reads the file and uploads it to the inference.sh cloud infrastructure.
Evidence: belt app run falai/topaz-image-upscaler --input '{"image": "/path/to/photo.jpg"}' as documented in SKILL.md and references/running-apps.md.
[INDIRECT_PROMPT_INJECTION]: The skill presents an attack surface where an agent, if influenced by malicious external data, could be directed to upload sensitive local files (e.g., SSH keys, environment files) to the external service using the belt tool's file upload capability.
Ingestion points: File paths provided to belt app run commands.
Boundary markers: None identified in the provided instructions.
Capability inventory: Execution of the belt CLI with network access and file read capabilities via Bash(belt *) tool.
Sanitization: None identified; the CLI automatically handles file resolution and upload.

agent-tools