AI Cluster Pre-flight Check

Pre-flight check for GPU cluster nodes — node validation before training, check cluster node health, is my GPU node ready. 26 health checks covering GPU, PCI...

This skill appears to do what it claims (local and cross‑node hardware/config checks), but take precautions before running it on production hosts: - Run it in a safe environment first (a non‑production node or an isolated VM) to observe behavior. - Be aware it may require root for full coverage (dmidecode, setpci, ipmitool, reading /dev/mem, etc.). Without root some checks will fail or be skipped. - Cross‑node checks (PREFLIGHT_PEER_IPS) will attempt SSH to the IPs you supply; those are outbound connections from the node. If you set SWITCH_HOST or similar it may attempt SSH to switches. - The script will run docker run with public vendor images (nvidia/cuda, rocm images). That will pull and execute container code from registries — review those image names and ensure pulling external images is acceptable in your environment. - The manifest under‑declares some runtime binaries (ssh, ip, lspci, ethtool, setpci, dmidecode, ipmitool, docker). Ensure required tooling is present and acceptable. - PREFLIGHT_NODE_ID is not a secret; the registry's labeling as a primary credential is misleading. If you want to proceed: review the bundled scripts (they are included), run with PREFLIGHT_PEER_IPS unset (local checks only) and without PREFLIGHT_STRICT first, and consider auditing or removing the docker tests if pulling/executing images is unacceptable in your environment.

Type: OpenClaw Skill Name: xperf-pre-flight Version: 1.0.1 The skill bundle contains several high-risk shell injection vulnerabilities in 'lib/checks.sh'. Specifically, 'check_1_25' uses 'eval' on the 'SWITCH_CLI_CMD' environment variable, and 'check_1_1', 'check_1_26', and 'check_mesh_ping' execute 'ssh' and 'ping' commands using unsanitized input from 'PREFLIGHT_PEER_IPS'. While these functions are aligned with the stated purpose of GPU cluster validation, the lack of input sanitization allows for arbitrary command execution if the environment variables are manipulated. No evidence of intentional malice or data exfiltration was observed.