The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted training data from external sources, which creates a surface for indirect prompt injection. 1. Ingestion points: Training scripts and recipes (e.g., sl_basic.py, sl_loop.py, rl_loop.py) load data from local JSONL files (via FromConversationFileBuilder) and Hugging Face datasets (via datasets.load_dataset). 2. Boundary markers: The skill uses specialized chat renderers (e.g., qwen3, llama3, tml_v0) that insert structural tokens like <|im_start|> and <|eot_id|> to delimit conversation turns. 3. Capability inventory: The skill possesses capabilities to make network API calls (tinker.ServiceClient), execute training passes (forward_backward), generate model output (sample), and write data to the local file system (save_state). 4. Sanitization: While data is processed by model-specific tokenizers, the skill lacks explicit mechanisms to validate or sanitize the natural language content for embedded malicious instructions.
[EXTERNAL_DOWNLOADS]: The skill performs multiple external retrieval operations. 1. It downloads training datasets from the Hugging Face hub (e.g., HuggingFaceH4/no_robots, openai/gsm8k, open-thoughts/OpenThoughts3-1.2M). 2. It fetches image content from external URLs in code examples using the requests library (e.g., example.com).
[COMMAND_EXECUTION]: The documentation provides instructions for installing dependencies using package managers, specifically 'pip install tinker tinker-cookbook' and variant installs for multimodal support.
[DATA_EXFILTRATION]: The skill performs network operations to external domains including thinkingmachines.ai (for documentation and API services) and example.com (for image data in recipes), which are not on the whitelisted domains list.

tinker