The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill uses torch.load to load pre-trained weights in scripts/utils/predictor/evaluation.py and scripts/kronos_predictor.py. In evaluation.py, the weights are loaded without the weights_only=True parameter, and in kronos_predictor.py, there is a fallback to legacy loading if safe loading fails. This facilitates arbitrary code execution via malicious .pt files (pickle-based exploitation) placed in the models directory.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted content from the web and news databases, which is then passed into LLM prompts for analysis and forecast adjustment.\n
Ingestion points: External news content and search results from Jina, DuckDuckGo, and Baidu.\n
Boundary markers: Prompts lack specific instructions or strong delimiters to prevent the LLM from following commands embedded in the external text.\n
Capability inventory: The skill allows the agent to modify forecasts and perform database writes based on the processed information.\n
Sanitization: The skill does not sanitize external text before interpolation into prompts.\n- [EXTERNAL_DOWNLOADS]: The skill is configured to automatically download machine learning models and tokenizers from Hugging Face (NeoQuasar/Kronos-*) and embedding models via sentence-transformers if they are not present in the local cache.\n- [SAFE]: The eval() findings in evaluation.py and training.py identified by static analysis are false positives. These refer to the model.eval() method in PyTorch, which is standard procedure for setting a model to evaluation mode, rather than the dangerous Python eval() function.

alphaear-predictor