ember

Security Policy

Reporting a vulnerability

Do not open a public GitHub issue for a security vulnerability.

Report vulnerabilities privately through GitHub’s private vulnerability reporting: open the repository’s Security tab and choose Report a vulnerability. Only you and the maintainers can see the report.

Include as much detail as you can: the affected component, a description of the issue, steps to reproduce, and your assessment of impact. A minimal proof-of-concept is helpful but not required for the initial report.

Response timeline

Stage Target
Acknowledgment Within 5 business days of receipt
Assessment and confirmation Within 10 business days
Coordinated disclosure Agreed with reporter before any public disclosure

If you haven’t received an acknowledgment within 5 business days, follow up via the same channel. Response times may be longer around public holidays.

Scope

In scope

The following areas are in scope for security reports:

Out of scope

Security-sensitive design notes

The HTTP server binds to 127.0.0.1 by default. A user may configure a remote inference endpoint (EMBER_SERVER_URL); the client refuses plaintext http to a non-loopback host unless EMBER_ALLOW_INSECURE_TRANSPORT is set, and a custom credential header is meant for a proxy that translates it. To serve other machines, set EMBER_HOST and set EMBER_SERVER_AUTH_TOKEN so /v1/systemone requires Authorization: Bearer <token> (compared with secrets.compare_digest); /health and /metrics stay unauthenticated for probes. Ember does not terminate TLS, so front remote-serving deployments with a proxy. When EMBER_SERVER_AUTH_TOKEN is unset the HTTP endpoint has no authentication layer and the security boundary is the network (loopback or an operator-managed proxy).

If you deploy ember in an environment where the loopback interface is shared (e.g., a multi-user server or a container with a shared network namespace), you are responsible for adding appropriate access controls. This configuration is not supported and is outside the intended threat model.

Hosted deployment (EMBER_MODEL_S3_URI) — operator trust model

When EMBER_MODEL_S3_URI is set, ember downloads a model snapshot from that S3 location at startup and imports joint_schema_model.py from the downloaded directory as executable Python. Ember performs no cryptographic integrity check on the downloaded content — this is an explicit architectural decision (constitution Article V, “Model Loading”). The security of a hosted deployment therefore depends entirely on the controls the operator applies to that S3 location:

The same principle applies to EMBER_MODEL_DIR: any directory pointed to by that variable is loaded and its joint_schema_model.py executed without verification. Keep its value out of user-supplied input.

Accepted residual risks (single-user loopback deployment)

The following are known, accepted risks for the default single-user loopback deployment. They are tracked in docs/stride-review.md and will not be fixed unless the deployment model changes.

Non-loopback binding without bearer auth (T-004): Setting EMBER_HOST to a non-loopback address (e.g. 0.0.0.0, required by platforms like Outerbounds that route traffic to a fixed container port) does not require EMBER_SERVER_AUTH_TOKEN to be set. /v1/systemone is unauthenticated in that configuration unless the operator sets the token or fronts the server with their own gateway/proxy auth — ember cannot distinguish “bound wide open on a hostile network” from “bound wide behind a platform’s own access-gated ingress” (e.g. Outerbounds’ auth.type: API, which authenticates every request before it reaches the pod). An earlier version of ember refused to start in this configuration at all; that refusal broke the documented Outerbounds deployment examples (deployment/flash.yaml, deployment/full.yaml), which deliberately relies on platform-level auth instead of ember’s own token. Setting EMBER_HOST to anything other than 127.0.0.1 without some access control — ember’s token, a reverse proxy, or a platform gateway — is the operator’s responsibility to get right. See vault/decisions/2026-10-10-remove-t004-non-loopback-auth-check.md.

Pidfile integrity (T-005): server.pid is a plain-text file with no cryptographic signature. The _is_ember_server check (verifying the process command string via /bin/ps) plus cross-checking the /health endpoint’s reported PID are strong compensating controls. An attacker would need local write access to the state directory and the ability to spawn a process with matching argv — a high bar on a single-user workstation.

/health and /metrics information disclosure (I-005, I-006): GET /health returns the server PID, package version, and whether bearer auth is required. GET /metrics exposes ember_model_info{model,device,dtype}. These endpoints are intentionally unauthenticated (they are used for health probes and Prometheus scraping). In the default loopback-only deployment any process on the same host can read them; in a shared-namespace environment (see above) they disclose server identity. If this is a concern, front the server with an authenticating reverse proxy.

No structured access log on /health / /metrics (R-005): uvicorn logs HTTP requests to server.log but in an unstructured format. For a single-user local tool this is acceptable. For shared environments, use a reverse proxy with structured access logging.

PID reuse TOCTOU (S-003): Between _is_ember_server confirming the PID and os.kill sending SIGTERM, the ember server could exit and the OS could reuse its PID for an unrelated process. The combined _is_ember_server + /health PID cross-check reduces this window to near-zero in practice. This is an OS-level race that cannot be fully closed in userspace without kernel support (e.g., pidfd). Documented here as an accepted residual risk.

Disclosure policy

This project follows coordinated disclosure. Once a fix is ready and tested, the maintainer will:

  1. Publish a patched release.
  2. Open a GitHub Security Advisory with full details.
  3. Credit the reporter (unless they prefer to remain anonymous).

Public disclosure happens after the patch is available, or after 90 days from the initial report if no fix is forthcoming, whichever comes first.