Skip to content

Security: inclusionAI/ASpec

Security

SECURITY.md

Security

ASpec is an early source release for trusted training environments, not a multi-tenant security boundary or a service intended for direct public-internet exposure.

  • Treat model/dataset/config input as code-bearing or resource-bearing. Use trust_remote_code only for sources/revisions you have reviewed and trust.
  • Load trusted checkpoints only. Exact resume includes Python/PyTorch serialization; a completion marker or matching hash does not make an untrusted checkpoint safe.
  • Restrict Ray, Mooncake endpoints and storage to trusted members. Authentication is not a promise of end-to-end encryption or safe public exposure.
  • Keep credentials, Ray authentication files, private prompts, weights and data out of issues/commits. Prefer dedicated venvs and least-privilege filesystem/network access.
  • Do not use host-wide process cleanup or change system drivers to troubleshoot an untrusted job.

Use the repository owner's private security channel if one is available. This file does not establish a contact address, response SLA or bounty program. If no private channel is advertised, request one without disclosing exploit details or credentials publicly. Ordinary install/correctness bugs can use the issue channel with redacted reproductions.

There aren't any published security advisories