The Agent Skills Directory

[DYNAMIC_EXECUTION]: The run.sh script utilizes python3 -c to run complex Python logic by interpolating shell variables directly into Python code strings. This creates a potential vector for code injection if task names or model types are maliciously crafted.
[COMMAND_EXECUTION]: The skill heavily relies on subprocess.run in bridge.py and shell-based orchestration in run.sh to trigger secondary skills and system commands for model lifecycle management.
[EXTERNAL_DOWNLOADS]: The workbench orchestrates remote training on RunPod GPUs, involving interaction with external cloud infrastructure, resource provisioning, and weight transfers.
[DATA_EXFILTRATION]: The skill reads from and manages sensitive local data stores, including ~/.pi/assistant/shadow.jsonl and model_registry.json, which contain interaction history and model configuration.
[PROMPT_INJECTION]: An indirect prompt injection surface is present in synthesis_eval.py, where the skill automatically extracts historical user inputs from shadow.jsonl to generate new evaluation prompts for the assistant, potentially re-injecting malicious instructions found in previous logs.

assistant-lab