The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied JSON files (proxy_kpi.json, benchmark_kpi.json, mining_pool.json) containing human-written prompts that are used in model training and evaluation.
Ingestion points: Data enters the agent context in scripts/validate_sharegpt.py and is subsequently interpolated into specs and synthetic record generation.
Boundary markers: While the data is structured (JSON), there are no specific markers or instructions to ignore embedded commands within the extracted prompt text.
Capability inventory: The skill possesses powerful capabilities, including arbitrary shell command execution via helper scripts and the ability to launch Docker containers with host-system mounts.
Sanitization: Data used in HTML report rendering is sanitized using html.escape in scripts/render_report.py to mitigate XSS risks.
[COMMAND_EXECUTION]: The workflow relies on executing local Python scripts and Docker containers to perform machine learning tasks.
Implementation: Scripts such as submit_cfw_train.py, submit_cfw_evaluate.py, and deft_exec.py use subprocess.run with list arguments, which is a secure implementation that prevents shell injection.
Policy Enforcement: scripts/deft_exec.py serves as a security wrapper that restricts the use of network and package management tools when an air-gap policy is active.
[EXTERNAL_DOWNLOADS]: The skill performs downloads from established external services as part of its setup and execution.
Sources: Models are fetched from Hugging Face and testing data is retrieved from Amazon S3.
Context: These operations are documented parts of the machine learning workflow and target trusted organizations or well-known cloud providers, which minimizes the associated risk.

tao-run-deft-aoi-cosmos3