The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The SKILL.md file provides an installation command curl -fsSL https://install.boltz.bio/boltz-api/install.sh | sh that downloads a shell script from a remote server and executes it immediately with shell privileges. This pattern allows for unverified arbitrary code execution on the host system.
[EXTERNAL_DOWNLOADS]: The skill downloads the boltz-api CLI installer from https://install.boltz.bio, which is an external source not included in the primary trusted registries.
[COMMAND_EXECUTION]: The skill facilitates the execution of the boltz-api CLI and a bundled shell script scripts/persist.sh. These tools are executed with arguments such as run names and file paths derived from user-provided inputs, which can be exploited if the inputs are not properly validated.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by processing untrusted user data to generate YAML payloads for the Boltz API CLI. Ingestion points: User prompts for designing protein or small-molecule structures (documented in references/examples.md). Boundary markers: Absent from the instructions and prompt templates. Capability inventory: The boltz-api CLI tool (which performs network operations and file system writes) and the scripts/persist.sh script (which performs file system operations). Sanitization: There is no evidence of sanitization or validation of user-provided content before it is interpolated into the payloads used by the CLI tool.

boltz