e2e-proxmox

Warn

Audited by Gen Agent Trust Hub on Oct 1, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONPRIVILEGE_ESCALATIONCREDENTIALS_UNSAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONOBFUSCATION
Full Analysis
  • [COMMAND_EXECUTION]: The proxmox-run.py script executes numerous shell commands on the local machine and a remote Proxmox node via SSH. Evidence includes the use of subprocess.run to call ssh, pct, git, and other system utilities for container management. User-provided arguments such as --template, --storage, and --bridge are interpolated into shell commands, though protected by shlex.quote.
  • [PRIVILEGE_ESCALATION]: The skill operates with high privileges and configures unrestricted access within the guest environment. Evidence includes the requirement for a root SSH destination for the Proxmox node to perform lifecycle operations and scripts/bootstrap.sh configuring nanoclaw ALL=(ALL) NOPASSWD: ALL in the guest's sudoers configuration.
  • [CREDENTIALS_UNSAFE]: The driver script reads a local API key and transmits it to the remote test container, where it remains stored. Evidence includes the Run.install method reading the file specified by --credential-file and uploading it to the LXC using self.upload. Test containers are retained after execution for troubleshooting, meaning credentials persist in the remote filesystem until manual cleanup.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes data from external sources and user input to drive high-capability shell operations. Ingestion points include Git repository URLs, commit references, and infrastructure configuration strings. The script employs shlex.quote for shell argument sanitization and uses random run IDs and marker verification to ensure commands are only executed against guest containers created by the current run.
  • [DYNAMIC_EXECUTION]: The skill generates and executes shell scripts at runtime to coordinate the installation process. Evidence includes proxmox-run.py dynamically creating run.sh and run-env.sh strings, uploading them to the LXC, and executing them via runuser.
  • [EXTERNAL_DOWNLOADS]: The skill fetches software packages and source code from remote network locations. Evidence includes scripts/bootstrap.sh performing apt-get operations to install dependencies from Debian repositories and proxmox-run.py executing git clone from user-specified or discovered repository URLs.
  • [OBFUSCATION]: The skill uses URL decoding to process configuration metadata returned from the Proxmox environment. Evidence includes the use of unquote() in confirm_guest() to handle Proxmox's encoding of the container description field.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Oct 1, 2026, 01:59 PM