vast-gpu

Pass

Audited by Gen Agent Trust Hub on Oct 9, 2026

Risk Level: SAFECOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill uses the Bash tool to execute local commands including vastai CLI operations, ssh for remote access, and rsync for file synchronization.
  • [REMOTE_CODE_EXECUTION]: The skill facilitates the execution of arbitrary code on remote GPU instances. It uses ssh to run shell commands and Python code as the root user on these external servers.
  • [DATA_EXFILTRATION]: The rsync synchronization logic includes a broad set of file types (*.json, *.yaml, *.yml, *.txt, *.sh). If the project directory contains sensitive configuration files or secrets in these formats, they will be uploaded to the remote third-party GPU host.
  • [INDIRECT_PROMPT_INJECTION]: The skill analyzes external data sources such as EXPERIMENT_PLAN.md and experiment scripts to automatically decide on GPU models and budget allocations. An attacker capable of modifying these project files could influence the skill to provision expensive hardware, leading to unexpected financial costs.
  • Ingestion points: refine-logs/EXPERIMENT_PLAN.md, project scripts (.py, .yaml, .json).
  • Boundary markers: None present; the skill treats the contents of these files as trusted requirements.
  • Capability inventory: vastai search offers, vastai create instance (financial impact).
  • Sanitization: None present.
  • [PRIVILEGE_ESCALATION]: The skill defaults to connecting to remote instances as the root user, which is the maximum privilege level on the target system.
  • [SECURITY_PRACTICE]: The skill employs ssh -o StrictHostKeyChecking=no, which disables host key verification. This increases the risk of Man-in-the-Middle (MITM) attacks during the connection to the remote GPU instance.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 9, 2026, 08:18 AM