complexa-slurm

Warn

Audited by Gen Agent Trust Hub on Jun 23, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSCREDENTIALS_UNSAFEDATA_EXFILTRATION
Full Analysis
  • [COMMAND_EXECUTION]: The skill executes local bash scripts and remote shell commands via SSH to manage cluster jobs. There is a potential command injection surface in scripts/cluster_preflight.sh where environment variables like CLUSTER_PARTITION are interpolated directly into remote SSH command strings without sufficient escaping (e.g., ssh ... "sinfo -p ${CV[CLUSTER_PARTITION]} -h").
  • [CREDENTIALS_UNSAFE]: The skill requires and manages access to a remote cluster via an SSH private key specified in the CLUSTER_SSH_KEY environment variable. It also uses insecure SSH options in its preflight script (-o UserKnownHostsFile=/dev/null), which bypasses host key verification and exposes the connection to Man-in-the-Middle (MitM) attacks.
  • [EXTERNAL_DOWNLOADS]: The skill is designed to download large external assets, including community models and checkpoints, from remote repositories and rsync results from user-defined cluster hosts.
  • [DATA_EXFILTRATION]: The skill performs bidirectional data transfer, rsyncing the local repository to a remote cluster and downloading processing results back to the local environment, which involves transmitting potentially sensitive code and data to external infrastructure.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Jun 23, 2026, 03:56 PM
Security Audit — agent-trust-hub — complexa-slurm