The outcome
Run a reproducible model job without exposing keys, datasets, management ports, or persistent sensitive data.
Step by step
A workflow you can repeat.
- 01
Define the job, GPU requirement, duration, region, model and data rights, threat model, network needs, storage, output, budget, and verified teardown criteria.
- 02
Create a task-specific SSH key, verify host details, restrict ingress, use a non-root account, patch the image, and never place long-lived credentials in shell history or repositories.
- 03
Transfer encrypted minimal data, verify checksums, pin containers and dependencies, isolate secrets, disable public services, and record the environment before execution.
- 04
Monitor GPU, process, network, disk, logs, spend, and unexpected access while validating outputs and preventing user-controlled code or prompts from reaching the host shell.
- 05
Export only approved results, delete data and credentials, terminate the rental, verify billing and loss of access, rotate exposed secrets, and retain a reproducible job manifest.
Working standard
What good use looks like.
- Use a task-specific SSH key.
- Expose no unnecessary ports.
- Verify deletion and termination.
Official references