The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The documentation explicitly recommends the command curl -fsSL https://cli.inference.sh | sh for installation. This is a high-severity pattern because it downloads and executes a remote script in the host's shell environment without prior verification or sandboxing.
[EXTERNAL_DOWNLOADS]: The skill relies on binaries and configuration files downloaded from cli.inference.sh and dist.inference.sh at runtime or during setup. While the skill claims to verify SHA-256 checksums, the initial bootstrap script remains a critical trust-link.
[COMMAND_EXECUTION]: The skill requests permission to use Bash to execute infsh commands. The security of these operations is entirely dependent on the integrity of the binary installed via the untrusted remote execution script.
[PROMPT_INJECTION]: The skill processes user-provided strings and external image URLs through the CLI, posing a risk for indirect prompt injection.
Ingestion points: Untrusted data enters via the prompt and reference_images fields in the infsh app run examples in SKILL.md.
Boundary markers: Absent. User input is interpolated directly into JSON structures for CLI execution.
Capability inventory: The skill uses Bash(infsh *) to perform network-based inference and local command execution.
Sanitization: No explicit sanitization or validation of input prompts or image metadata is implemented in the skill definition.

qwen-image