Do not disclose suspected vulnerabilities, exploit instructions, credentials, or private operational details in public GitHub issues or pull requests.
Report vulnerabilities privately: use Report a vulnerability on this repository's Security tab where GitHub offers it, or send the report to the current maintainer, Wenyi Wang, at wenyiw@uchicago.edu. Do not use a public issue as a substitute for a private vulnerability report.
In a private report, include:
- A description of the issue and its potential impact.
- The exact source commit or package version and artifact hash, if known.
- The affected configuration, site/profile, scheduler, vendor runtime, engine, gateway, and relevant dependency versions.
- Minimal reproduction steps and sanitized evidence; share only what is needed.
- Any known mitigation or proposed fix.
Do not send live credentials, access tokens, private keys, personal information, restricted datasets, or sensitive prompts. If credentials may be compromised, follow your organization's credential-revocation and incident-response process. Reports affecting a shared HPC system should also follow the site's security reporting procedure; this project cannot replace the site's incident response.
ExaServe is under active production hardening. There is no published long-term security-support or backport matrix. Report suspected vulnerabilities against any identifiable revision; maintainers must determine affected revisions and the appropriate remediation scope. Do not infer support from a version number, the existence of a development branch, or a passing CI check.
The README and hardening status describe the intended runtime boundary, exact candidates, superseded evidence, and open qualification gates. Qualification is specific to the artifact and deployment dimensions; it is not a guarantee that a deployment is free of vulnerabilities. The current production-candidate design assumes a trusted allocation-internal network and is not a public Internet service security boundary.
ExaServe's inference endpoints have no authentication, TLS or rate limiting. HAProxy and the per-node Ray Serve proxies listen on all interfaces of the allocated nodes, and Ray's own ports are unauthenticated in this configuration. Anyone who can connect to those nodes can send inference requests, and anyone who can reach the Ray ports can run code as the job user. Deploy ExaServe only where the site network limits these ports to users you trust; ExaServe does not check this.
Model weights, config.json, tokenizer and chat-template files are trusted
input and are treated like code. ExaServe loads them with
trust_remote_code=False, Hugging Face offline mode and an architecture
allowlist (LlamaForCausalLM, GptOssForCausalLM) that also pins each
architecture's model_type, but these checks are not a sandbox. Stage only models from sources you trust, and keep
model_storage_path writable only by the operator. Requests cannot supply
their own chat template.
Ray, vLLM, PyTorch and Transformers come from the qualified Aurora
frameworks/2025.3.1 module, loaded from the 26.26.0 image's module tree (see
getting started and
requirements/aurora-frameworks-2025.3.1.lock),
and change only through requalification. Advisories against those versions are
assessed against the request surface in
docs/api.md. Low-severity engine defects
that ordinary requests can reach are accepted until the next requalified stack.
Maintainers should coordinate investigation, fixes, and disclosure privately, then publish appropriate remediation guidance and security advisories after assessing exposure and affected users. No response-time or fix-time service level is currently promised. Avoid public disclosure of actionable details while a private report is being assessed; agree on disclosure timing with the maintainers and affected operators.