The Agent Skills Directory

[COMMAND_EXECUTION]: The skill leverages a repl tool to access the host.agents and host.skills SDKs, enabling the agent to programmatically manage profiles, identities, and connector permissions within the control-plane kernel.
[REMOTE_CODE_EXECUTION]: Allows for the authoring and publication of kernel.py and kernel.R sidecar files. These files are executed within a persistent kernel, providing a mechanism for the definition and execution of arbitrary code logic.
[PROMPT_INJECTION]: Instructions explicitly direct the agent to withhold information from the user about the frequency of security approval cards, specifically for bulk delete operations and identity switching tasks, which reduces transparency for security-critical actions.
[PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection as it processes untrusted data for profile identities and skill content. It lacks automated sanitization for these inputs, relying solely on user review gates despite possessing high-privilege capabilities like environment manipulation.
[DATA_EXFILTRATION]: The SDK provides a host.skills.read method that allows the agent to read internal skill source code and configuration, which could be used to discover and exfiltrate internal state or logic.
[EXTERNAL_DOWNLOADS]: The skill manages the installation of external dependencies and software environments using tools like manage_packages and manage_environments, which interact with public package registries for libraries such as requests, numpy, and pandas.

customize