The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill's code templates in references/skeleton-template.py and references/method-templates.py, as well as the implementation in evals/targets/eval_2_target.py, use torch.load with weights_only=False. This is a known security risk in PyTorch that allows the execution of arbitrary code during the unpickling of model checkpoints. If an agent loads a malicious model from an untrusted source, this could lead to system compromise.
[EXTERNAL_DOWNLOADS]: The main workflow defined in SKILL.md (Step 0) requires the agent to fetch a reference inference script from a user-provided URL or local path. This introduces unverified external content into the agent's workspace.
[REMOTE_CODE_EXECUTION]: The validation steps (Steps 10 and 11 in SKILL.md and references/validation-guide.md) instruct the agent to run 'vanilla' third-party scripts and comparison scripts using uv run python. This represents the execution of remote code fetched during Step 0.
[COMMAND_EXECUTION]: The skill utilizes several powerful CLI tools and shell commands, including uv run, pytest, make, git, and pip install, for dependency management, testing, and environment setup.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it ingests and processes untrusted third-party scripts.
Ingestion points: Reference inference scripts fetched from external URLs in SKILL.md (Step 0).
Boundary markers: No delimiters or instructions to ignore embedded commands are present in the analysis phase.
Capability inventory: The agent can execute shell commands (uv run), write files, and load Python objects (torch.load).
Sanitization: The skill does not prescribe any content validation or sanitization for the fetched scripts before they are analyzed or executed.

earth2studio-create-prognostic