The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill facilitates the installation of external binaries and packages from GitHub and NPM registries using go install github.com/mvanhorn/printing-press-library/... and npx -y @mvanhorn/printing-press-library. These are vendor-owned resources but represent external code dependencies executed at runtime.
[REMOTE_CODE_EXECUTION]: The skill downloads and executes a binary (vehicle-safety-pp-cli) from a remote repository. It provides instructions to verify and run this compiled code directly within the agent's environment.
[DATA_EXFILTRATION]: The tool includes a --deliver webhook:<url> feature that allows sending the results of any command to an arbitrary HTTP endpoint. This capability could be abused by a malicious prompt to exfiltrate vehicle data, VIN details, or locally stored credentials to an attacker-controlled server.
[COMMAND_EXECUTION]: The skill relies heavily on the Bash tool to perform complex operations, including installing software, setting environment variables, and managing local state files like credentials.toml and data.db.
[DYNAMIC_EXECUTION]: The 'Learning Loop' and 'Playbook' mechanisms represent a dynamic execution surface. The agent is instructed to 'teach' the CLI new mappings and 'replay' steps from a playbook. If the data used during the 'teach' phase is poisoned, the agent may unknowingly execute malicious commands stored in the local playbook database.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user queries via the recall and teach commands to build a local knowledge base. Since these learned entries are used to resolve future commands and playbooks, there is a risk of the agent being manipulated through specially crafted inputs that are later re-executed as 'trusted' learnings.

pp-vehicle-safety