The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a bundled shell script scripts/compass_status.sh to perform environment diagnostics and state reporting.
[REMOTE_CODE_EXECUTION]: The diagnostic script uses python -c to dynamically execute Python code for initializing Isaac Sim and validating PyTorch checkpoints.
[COMMAND_EXECUTION]: A potential command injection vulnerability exists in scripts/compass_status.sh. The --ckpt argument is interpolated directly into a Python string literal inside a shell command: python -c "... torch.load('${CKPT_PATH}', ...)". An attacker who can control the file path could potentially execute arbitrary Python or shell commands if the file name is crafted to break out of the string literal, although this is mitigated by a preceding file existence check ([ -f "${CKPT_PATH}" ]).
[COMMAND_EXECUTION]: The skill instructions suggest using the dangerouslyDisableSandbox: true flag. This configuration allows scripts to escape the execution sandbox to access host resources like nvidia-smi for GPU diagnostics, which increases the security impact of any other vulnerabilities.
[DATA_EXFILTRATION]: The skill identifies and reports the presence of sensitive environment variables such as WANDB_API_KEY and HF_TOKEN. While it currently only checks for their existence using test -n, this logic establishes a pattern for accessing credentials.

compass-doctor