The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill explicitly enables the execution of remote code when loading tokenizers from the Hugging Face Hub.
Evidence: Found in scripts/mock_tokenizer.py, scripts/to_datum.py, and scripts/test_finetuned.py where AutoTokenizer.from_pretrained is called with trust_remote_code=True.
Risk: This setting allows the execution of arbitrary Python code defined in the model's repository, which could lead to complete system compromise if the repository is malicious or compromised.
[DYNAMIC_EXECUTION]: The skill uses the transformers library to dynamically load and execute model-specific code at runtime.
Evidence: Multiple occurrences of trust_remote_code=True across the script files.
[DATA_EXPOSURE]: The skill contains hardcoded filesystem paths that reveal internal directory structures.
Evidence: scripts/train_lora.py defines a default data path at C:\Users\admin\Downloads\pen200_datum_training.jsonl.
Risk: Hardcoded paths can expose sensitive information about the host system's user accounts and environment.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process data from external PDF files which serves as a potential attack surface for indirect prompt injection.
Ingestion points: scripts/chunk_pdf.py extracts raw text and OCR-processed text from PDF pages (lines 78 and 93).
Boundary markers: No delimiters or specific instructions are provided to the agent to treat the extracted PDF content as untrusted data.
Capability inventory: The skill has file-write access (open(..., 'w')), network capabilities via the Tinker SDK, and the ability to execute remote code via trust_remote_code.
Sanitization: The extracted content is processed and saved to training files without any sanitization, filtering, or escaping of potential malicious instructions embedded in the document text or images.
[CREDENTIALS_UNSAFE]: The skill includes instructions and code for managing API keys that could lead to accidental exposure.
Evidence: scripts/test_finetuned.py and scripts/train_lora.py include logic to hardcode TINKER_API_KEY if the environment variable is missing.

pdf-to-tinker-training