Security fixes target the current main branch and the latest tagged minor
release. Older release lines receive fixes only when explicitly announced.
Please report security issues privately to Sonic Field Labs before public disclosure. Include the affected commit, local configuration, reproduction steps, and whether raw audio or private listening traces can be exposed.
oida binds to 127.0.0.1 by default. Wildcard or LAN binds are refused unless
OIDA_AUTH_TOKEN (legacy HMM_AUTH_TOKEN/AEAR_AUTH_TOKEN) is set, and
token-protected clients must send Authorization: Bearer <token>. Loopback
requests are additionally guarded by Host/Origin checks against DNS rebinding
and cross-origin calls.
Audio-analysis routes intentionally accept an operator-selected local file path; these are desktop/localhost file references, not web-root-relative tenant paths. The daemon verifies that each reference resolves to an existing regular file before analysis. Writes remain confined to configured Oída data directories, and names used inside those directories are normalized. The Sonic Field reveal route separately requires the resolved target to remain inside its configured Sonic Field root.
MOSS-Audio Hugging Face model lookup is disabled by default. Download weights
into weights/ or set OIDA_ALLOW_HF_HUB=1 (legacy HMM_/AEAR_) explicitly.
HF_HUB_OFFLINE=1 always disables hub lookup.
The optional local runtime retains Torch and Torchaudio 2.10.0. Updating this pair requires separate model, dependency and accelerator qualification; this review does not claim new GPU/MPS or model-inference evidence.
The machine-readable policy is advisory-exceptions.json. The repository CODEOWNER, @emeisazam, owns follow-up. Reviewed 26 September 2026; exceptions expire 10 October 2026, or earlier if the affected API, model trust boundary or runtime version changes. Review again before enabling an optional runtime. CI refuses an expired policy or a widened advisory set before invoking pip-audit; it suppresses exactly these two advisory identities.
| Advisory | Current upstream evidence | Scoped repository review |
|---|---|---|
PYSEC-2026-139 / CVE-2026-4538 |
.pt2 deserialization advisory; no patched version is listed by the current audit feed. | No direct torch.export.load call or .pt2 model intake found in owner source. |
CVE-2025-3000 / GHSA-rrmf-rvhw-rf47 |
TorchScript advisory; fixed in 2.13.0. | No direct torch.jit.script call found in owner source. |
This is a bounded source review, not a transitive execution trace or permission to load untrusted model artifacts. The reviewed optional runtime remains conditional on its existing operational qualification requirements.
The September review also found two AnyIO advisories in the locked 4.13.0: TLS hostname handling and process-pool stderr handling. Both have fixes in 4.14.2. The dependency floor now excludes older versions; these findings are not suppressed. New audit findings continue to fail CI.
Repository call-site inspection found no new use of the excepted APIs. The all-extras dependency audit passes with exactly the same two exceptions; no additional advisory was suppressed. HTTPX2/HTTPCore2 were updated to 2.12.0, and the affected optional dependency bounds were refreshed. These software checks do not validate GPU/MPS inference with the new resolved dependencies. Keep optional model execution conditional on that validation and review the exceptions again before enabling it, or by the deadline above.