autoresearch-ml

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.80). The skill's prepare.py explicitly downloads public parquet shards from HuggingFace (BASE_URL pointing to huggingface.co/datasets/karpathy/climbmix-400b-shuffle) and ingests their text into the tokenizer and training dataloader used by the autonomous experiment loop, so arbitrary user-generated content from the open web is consumed at runtime and can materially affect training outcomes and experiment decisions.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.80). The pyproject specifies an external package index (https://download.pytorch.org/whl/cu128) which is used during the skill's setup step ("uv sync") to fetch/install PyTorch wheels (remote code that will be executed/loaded) and the skill requires these packages to run.