Skip to content

v0.8.0

Choose a tag to compare

@github-actions github-actions released this 14 Nov 06:01
· 56 commits to main since this release
c3fd272

Guard Spec v0.8.0 puts the gateway in front of an organization's own JSON APIs, not only LLM providers, and makes redaction of personal data in answers a per-organization choice. Detection defaults are unchanged: the classifier stays v3b. Full list in the CHANGELOG. Documentation: https://guardspec.io/docs/.

Highlights

  • Custom backends. A new upstream kind, custom, for any JSON API of your own, reached through /v1/backend/{upstream}/{path} with the usual API key.
    • You declare the allowed path prefixes and methods, and the JSONPaths of the request fields to scan and of the answer fields to guard (for example $.query, $.messages[*].content).
    • Those fields and every query value go through the same detectors, thresholds and policies as chat. A refused request never reaches the backend; an allowed one is forwarded byte for byte.
    • The answer's declared fields get the output guard with the organization's personal data settings. Event streams pass unchecked.
    • The backend key stays on the gateway (api_key_env under UPSTREAM_KEY_ORGS); client credentials and hop-by-hop headers are not forwarded, and redirects are not followed.
    • Admin UI: Model providers → Custom backend. Example: examples/custom-backend. Docs: integrations.
  • Personal data in answers, per organization.
    • Guard settings output_pii_action (off or redact) and output_pii_kinds (email, phone, card, cccd, ipv4) override OUTPUT_PII_REDACT and OUTPUT_PII_IPV4 for one organization, on every chat API, streamed or not.
    • Admin UI: Guard settings → Personal data in responses. Left unset, nothing changes.
    • Values replaced by input_pii_action: tokenize still come back in the answer.
  • Research: R7 did not meet its target, no new model. A pre-registered training round paired benign long text with attacks embedded in the same documents. On a new holdout scored once, the candidate blocked 2 of 180 clean long documents (the released v6: 10) and caught 95.1% of embedded attacks, but lost precision on short prompts, so it is not released. Reports: train/reports/v8_holdout.md, MODEL_CARD "R7".
  • Base image. The gateway images are built on a newer Chainguard glibc-dynamic (built 2026-11-12).

Upgrade notes

  • Apply migration 031 before a 0.8.0 gateway starts. It adds the provider custom and the column upstreams.backend, which the gateway reads with every upstream. The compose migrate job does this. Existing upstreams are not changed; the down migration deletes custom backends.
  • No behaviour change by default. output_pii_action and output_pii_kinds are unset after the upgrade, so every organization keeps OUTPUT_PII_REDACT and OUTPUT_PII_IPV4 until an admin sets them.
  • See docs/upgrade.md.

Verify

cosign verify ghcr.io/guard-spec/guard-spec-gateway:0.8.0 \
  --certificate-identity https://github.com/guard-spec/guard-spec/.github/workflows/release.yml@refs/tags/v0.8.0 \
  --certificate-oidc-issuer https://token.actions.githubusercontent.com

gh attestation verify guard-spec_0.8.0_linux_amd64.tar.gz --repo guard-spec/guard-spec

Which build runs which detection layers

Artifact Layers Notes
ghcr.io/guard-spec/guard-spec-gateway:0.8.0 L1 + L2 + L3 Built with -tags onnx, ships libonnxruntime. Mount the models at /models (read-only is fine); without them it runs L1 only and logs a warning. GET /readyz lists the active layers.
ghcr.io/guard-spec/guard-spec-gateway-full:0.8.0 L1 + L2 + L3 The gateway image with guardspec-classifier-v3b and multilingual-e5-small baked in (read-only), for air-gapped installs. Built only when the GATEWAY_FULL_L2_SOURCE variable is set.
guard-spec_0.8.0_linux_<arch>_onnx.tar.gz L1 + L2 + L3 Gateway built with -tags onnx plus libonnxruntime.so. Linux, glibc 2.36 or newer; set ONNXRUNTIME_LIB and MODEL_DIR.
guard-spec_0.8.0_<os>_<arch>.tar.gz L1 only Static binaries (CGO_ENABLED=0), no ONNX Runtime.
ghcr.io/guard-spec/guard-spec-control-plane:0.8.0 n/a Static, no cgo. Serves the admin UI at /ui/.

Apart from guard-spec-gateway-full, models are never baked into images or archives; fetch them with models pull (see README).

Changelog

See the CHANGELOG and the full diff.