Security fixes target the current main branch and the latest public release.
Older releases may receive fixes at maintainer discretion. Please confirm a
finding against current main where you can.
Please do not open a public issue or pull request for a suspected vulnerability, and please do not disclose details publicly until a fix is available.
Report privately through either channel:
- GitHub: open the Security tab and choose Report a vulnerability. The report stays visible only to you and the maintainers.
- Email: send the report to support@liquid.ai.
Maintainers should acknowledge reports within 7 business days and coordinate next steps with the reporter.
A useful report includes:
- what an attacker can do, and what they need in order to do it (network reachability, a registered client identity, an operator's storage credentials);
- the affected version or commit;
- steps to reproduce, ideally against a local
serveinstance; - any log output or request/response capture that shows the behavior, with keys, signatures, and tokens redacted.
This repository is the management server: the HTTP API, submission verification, scoring orchestration, and warehouse writes. Findings in the client harnesses belong to pipette-clients.
The security-relevant surfaces here are:
- request authentication and identity: device signatures, user sign-in and
keys, and the permission model
(
docs/authentication.md); - the vault, which holds customers' private model weights and serves them only to their owner's devices (Vault);
- the submission path that accepts and stores client-supplied data;
- the PostgreSQL database that holds the records (accounts, organizations, devices, plans, campaigns, vault records), and object-store access.
Reports we are glad to receive but do not treat as vulnerabilities in this project:
- findings that require an operator's own database or object-store credentials, or filesystem access. That trust is assumed by the deployment model;
- vulnerabilities in third-party crates, unless this project's use of the crate is what makes them exploitable. Report those upstream; tell us as well if a version bump here is the fix.