The Agent Skills Directory

[DYNAMIC_EXECUTION]
Use of shell eval and unsafe pickle loading.
Found in scripts/run_one.sh.template and run_queue.sh.template: The templates use the eval command to execute the PROXY_HOOK variable which contains platform-specific setup commands.
Found in scripts/verify_local.py: The script allows unsafe pickle deserialization via torch.load when the --allow-pickle flag is provided. It includes prominent warnings about the risks.
[INDIRECT_PROMPT_INJECTION]
Data ingestion surface from remote logs.
Found in scripts/health_patrol.sh.template: The script performs grepping on remote log files and echoes values that the agent parses to determine job status. This provides a surface where malicious content on a remote machine could influence agent logic.
[COMMAND_EXECUTION]
Use of destructive process termination.
Found in scripts/reap_vram_zombies.sh: The script uses kill -9 to terminate processes identified as GPU zombies. Its use is restricted by a --force flag.
[EXTERNAL_DOWNLOADS]
References to external package and model registries.
The skill references resources from pypi.org, huggingface.co, aliyun.com, and tuna.tsinghua.edu.cn for downloading models and dependencies. These references are to well-known services and mirrors.

remote-gpu-trainer