The Agent Skills Directory

[COMMAND_EXECUTION]: The shell scripts scripts/run_one.sh.template and scripts/run_queue.sh.template utilize the eval command on the PROXY_HOOK variable. This variable is intended to be populated by the agent based on platform-specific profile documentation. This pattern allows for arbitrary command execution on the local machine if a profile contains malicious shell commands.
[REMOTE_CODE_EXECUTION]: In scripts/verify_local.py, the script allows loading PyTorch checkpoints with weights_only=False. This enables the execution of arbitrary code via the Python pickle module. Although this is gated behind the --allow-pickle flag, it presents a significant risk if an attacker can provide a compromised checkpoint file for verification.
[CREDENTIALS_UNSAFE]: The skill's design involves reading sensitive API keys from local files and environment variables. For example, scripts/run_one.sh.template uses cat to read credentials from a file path provided in the CRED_FILE variable, which can lead to accidental exposure or logging of secrets in multi-tenant environments.
[COMMAND_EXECUTION]: The script scripts/reap_vram_zombies.sh can execute kill -9 on system processes. It identifies targets by scanning /dev/nvidia* and checking process visibility in nvidia-smi. This capability could be abused to terminate legitimate system processes if the logic is manipulated or if targeting criteria match unintended processes.

remote-gpu-trainer