The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads and installs a Python wheel from a remote domain not included in the trusted vendor list.
Evidence: pip install ... "tuplatform-connect @ https://connect.aiscientist.tools/downloads/tuplatform_connect-0.3.0-py3-none-any.whl#sha256=3fad5eee5ecf7887a693d93ccd1aa112dc0955617a885d1fc3daded0030f9ae0" in SKILL.md.
Note: The installation uses a SHA-256 hash pin, which mitigates tampering risks but the source remains an external third-party domain.
[INDIRECT_PROMPT_INJECTION]: The skill implements a remote tool that ingests untrusted sequence data, which could potentially contain malicious instructions targeting the agent.
Ingestion points: The run_borzoi_predict tool accepts a sequence parameter from the user/environment.
Boundary markers: The provided examples and instructions do not include specific delimiters or warnings to ignore instructions embedded in the sequence data.
Capability inventory: The skill utilizes subprocess execution via the tu CLI (tu serve, tu remote share) and starts a local HTTP server to host the MCP tool.
Sanitization: Instructions mention checking for scientific meaning and bounded values, but no explicit sanitization or escaping of the input sequence is performed within the skill logic.
[COMMAND_EXECUTION]: The skill executes multiple shell commands to manage environments, install dependencies, and launch remote services.
Evidence: Commands such as python -m pip install, tu remote login, and python -m tooluniverse.remote.borzoi.borzoi_tool are used throughout SKILL.md.

setup-borzoi-remote-tool