The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is specifically designed to execute arbitrary code. It utilizes E2B sandboxes to provide isolation by running code in fresh cloud VMs, which is a recommended security practice for handling untrusted AI-generated output.
[EXTERNAL_DOWNLOADS]: The documentation provides instructions to install official packages (e2b-code-interpreter and @e2b/code-interpreter) from standard package registries (PyPI and npm).
[CREDENTIALS_UNSAFE]: Uses a standard placeholder (your_key_here) for the E2B_API_KEY environment variable. This is a common instructional pattern and does not expose actual secrets.
[PROMPT_INJECTION]: The skill facilitates the execution of potentially untrusted code, creating an indirect prompt injection surface.
Ingestion points: Untrusted code enters the system via the run_code method in SKILL.md examples.
Boundary markers: Absent; the code string is passed directly to the execution function.
Capability inventory: The run_code method allows full code execution within the sandbox environment.
Sanitization: No input sanitization is present in the examples; the primary security control is the VM-level isolation provided by the E2B platform itself.

e2b-sandboxes