The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill interacts with sensitive files including SSH private keys (~/.ssh/id_ed25519) and authentication tokens (~/.config/enroot/.credentials). These files are accessed and modified to enable non-interactive remote access to computing clusters.
[COMMAND_EXECUTION]: Multiple system utilities such as ssh, sbatch, srun, and scancel are invoked using environment variables like SLURM_USER and SLURM_HOSTNAME. This pattern can facilitate command injection if the variables are not correctly sanitized before being passed to the shell.
[EXTERNAL_DOWNLOADS]: The skill instructs the user to install the nvidia-tao-sdk package from the public Python Package Index (PyPI). This is a standard operation for skills designed to work with NVIDIA toolkits and originates from a well-known service provider.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it processes untrusted data, such as Lustre filesystem paths and cluster configurations, which are then used in command execution.
Ingestion points: Environment variables (SLURM_USER, SLURM_HOSTNAME) and dataset URIs (lustre:///) provided by the user.
Boundary markers: No specific delimiters or "ignore" instructions are used when interpolating these values into shell commands.
Capability inventory: Includes remote shell execution via ssh and cluster job management via sbatch.
Sanitization: The provided instructions do not include mechanisms for validating or escaping the contents of environment variables before they are used in scripts.

tao-run-on-slurm