# Changelog

## 0.3.0 (2026-11-06)

Stricter verdicts on the checks of 0.2.4. Online, the verdict and the exit code are those of
0.2.4, with one more warning when the TD measurements are not pinned. Offline, the limits 0.2.4
printed as warnings are now part of the verdict, with exit code 3, or required checks. Upgrade
with `pip install -U voltage-verify`; scripts that run `verify --offline` or `quote --offline`
must handle exit code 3 (see Compatibility).

What changes from 0.2.4:

* Offline, the TCB status and the revocation lists come from the bundle's Intel collateral.
  0.2.4 said so in the warning `tdx.tcb_offline`; the verdict now says it, `RESULT: VERIFIED
  (TCB and revocation from the bundle's Intel collateral, not checked against Intel today)`,
  with exit code 3. `quote --offline`, which checks no TCB status and no revocation, ends on
  `RESULT: VERIFIED (signatures only, TCB not checked offline)` and exit code 3, where 0.2.4
  ended on `RESULT: VERIFIED` and exit code 0. `--accept-offline-tcb` (`verify`, `quote`)
  accepts the limit: plain `RESULT: VERIFIED`, exit code 0. `tdx.tcb_offline` is gone.
* Offline, `tdx.collateral_fresh` is a required check. It fails when the oldest issue date Intel
  signed into the TCB info, the QE identity and the PCK CRL is more than 30 days old (the
  validity Intel gives these documents), when the root CA CRL is past its nextUpdate, or when
  the CRLs are missing. 0.2.4 printed a warning starting with `STALE:`.
  `--max-collateral-age DAYS` (`verify`, `selftest`) accepts older collateral explicitly. Online
  behaviour is unchanged.
* Without `--expect-mrtd` (and `--expect-rtmr0` to `--expect-rtmr3`), `tdx.measurements` is a
  warning, `TD software not pinned: pass --expect-mrtd to check it`, where 0.2.4 printed no
  line. `VERIFIED` is unchanged.
* The dates of the root CA CRL are decoded as strict base64; no change for a well-formed bundle.

Compatibility:

* Exit codes: 0 verified, 1 a required check failed, 2 unusable input, 3 verified with a limit
  named in the verdict (offline TCB). A script that treats anything but 0 as a failure fails
  offline runs until it passes `--accept-offline-tcb` or verifies online.
* `--json`: new keys `verdict` and `limits`; `ok`, `checks` and `facts` are unchanged. Offline,
  `tdx.collateral_fresh` is required and `tdx.tcb_offline` no longer appears. A script that
  treats any warning as a failure fails on every run that does not pin the TD measurements.
* Library: `verify_bundle` and `verify_quote` take `accept_offline_tcb`. New: `Report.limits`,
  `Report.verdict`.

## 0.2.4 (2026-11-05)

Security release for both sides of the verifier, from an audit of 5 November 2026, with no
change for scripts: the `RESULT` line, the exit codes and the `--json` keys are those of 0.2.3,
and the eight bundles VoltageGPU published from 12 September to 4 November 2026 still verify,
online and offline, with exit code 0. Upgrade with `pip install -U voltage-verify`.

What 0.2.3 and earlier accepted, and what is refused now:

* NVIDIA GPU tokens taken from another attestation. NRAS lists every GPU token in the verifier
  token (`submods`, the SHA-256 of the compact token) and each GPU token carries `eat_nonce`;
  neither was checked, so a GPU token from another NVIDIA attestation verified next to a fresh
  verifier token and could pad `--gpus`. The GPU tokens must now be exactly the ones the
  verifier token lists (`nvidia.signatures`) and carry the manifest's nonce (`nvidia.nonce`).
  `selftest` adds `nras-extra-gpu` (ten mutations).
* Offline, Intel collateral with a second copy after its signature. A TCB info or QE identity
  body `{"tcbInfo":<signed>,"signature":"...","tcbInfo":<forged>}` verified, and the forged copy
  was the one read: a bundle could carry any TCB status and any issue date. The body must now
  be exactly the signed document followed by its signature, and is read from the signed bytes.
  Collateral fetched online from Intel was not exposed.
* Online, every intermediate NVIDIA's live JWKS presented became a trust anchor. Only the pinned
  GPU Intermediate 004 and 005 are now, online and offline, and the live JWKS is not fetched;
  `verify --trust-live-intermediates` restores the previous behaviour after a rotation by
  NVIDIA. All published bundles chain to 004.
* A PCK chain whose issuers were not CAs. Every issuer must now carry BasicConstraints with
  `ca=True`.
* Malformed input stopped the verifier with a traceback, and an unreadable bundle printed no JSON
  with `--json`. An unusable bundle is now the failed check `bundle.malformed` (exit code 2,
  valid `--json` report); a malformed field fails the check that owns it. 0.2.3 verified none of
  these inputs.

What is said now, as warnings that change neither the verdict nor the exit code:

* Offline, the TCB status and the revocation lists come from the Intel collateral the bundle
  carries. `tdx.tcb_offline` says so (`TCB and revocation from the bundle's Intel collateral,
  not checked against Intel today`), and `tdx.collateral_fresh` gives the age of that
  collateral, the oldest issue date Intel signed into the TCB info, the QE identity and the PCK
  CRL, starting with `STALE:` past the 30 days Intel gives these documents. 0.2.3 said nothing
  until the collateral was past its nextUpdate. `--max-collateral-age DAYS` (`verify`,
  `selftest`) makes the age a required check, failed past DAYS.
* Without `--challenge`, `manifest.challenge`: `no challenge supplied: freshness and replay not
  checked`.
* On a multi-GPU mode, `nvidia.scope`: the per-GPU tokens do not attest the fabric between the
  GPUs (the NVSwitch fabric in Protected PCIe mode, the encrypted NVLink fabric in NVLE mode).
* `tdx.measured_compose`, when the bundle's environment carries a `dstack` block: the TD runs a
  composition measured by the provider (its compose hash lands in RTMR3), which the tool
  compares to no published reference: trusted, not proven.

New:

* `--expect-mrtd` and `--expect-rtmr0` to `--expect-rtmr3` (`verify`, `quote`) pin the TD's
  measurements, checked by `tdx.measurements`. Without them nothing changes and no line is
  printed; `docs/WHAT_IT_PROVES.md` says what an unpinned `VERIFIED` does not prove, and that
  NVIDIA's tokens name no TD.
* `nvidia.mode`: the NVIDIA attestation mode recorded in the bundle must be one the tool knows
  (`single-gpu`, `multi-gpu-ppcie`, `multi-gpu-nvle`); 0.2.3 did not check it.
* `attest --mode multi-gpu-nvle`, for a Blackwell node in NVIDIA NVLink-encryption (NVLE) mode:
  each GPU is attested on its own, through the same standalone path as `single-gpu`.
* `contrib/dstack-b200/`: a collector for a confidential GPU container run by dstack, where the
  tenant has no configfs TSM. It asks the dstack guest agent for a quote over the manifest's
  SHA-512, runs NRAS for every GPU and writes the same `voltage-verify-bundle/1`, with the
  `dstack` block (app id, event log) in its environment. It refuses to write a bundle unless the
  quote's report_data is the manifest hash byte for byte. Not installed with the package.

Compatibility:

* Unchanged: the last line (`RESULT: VERIFIED`, or `RESULT: NOT VERIFIED (<failed checks>)`),
  the exit codes (0 verified, 1 a required check failed, 2 unusable input) and the `--json` keys
  `ok`, `checks` and `facts`.
* New lines: the warnings above (`tdx.tcb_offline` and `tdx.collateral_fresh` on every offline
  run, `nvidia.scope` on every multi-GPU bundle), `nvidia.mode` on every run, `tdx.measurements`
  when pinned, `bundle.malformed` for an unusable bundle; new detail texts for
  `nvidia.signatures` and `nvidia.nonce`. A script that treats any warning as a failure fails on
  every offline run and every multi-GPU bundle.
* Online, NVIDIA's live JWKS is no longer fetched by default: a bundle signed under an NRAS
  intermediate this release does not pin fails until `--trust-live-intermediates` or a release
  that pins it.
* Library: `verify_bundle` takes `expected_measurements`, `max_collateral_age` and
  `trust_live_intermediates`, `verify_quote` takes `expected_measurements`;
  `verify_bundle(online=True)` no longer uses NVIDIA's live intermediates by default. New:
  `NrasResult.nonce_mismatch`, `pcs.signed_dates`, `pcs.COLLATERAL_VALIDITY`.
  `pcs.SignedDocument.payload` is rebuilt from the signed bytes.
* Dependencies: `cryptography>=43.0.1`, whose earlier wheels ship a vulnerable OpenSSL
  (GHSA-h4gh-qq45-vh27).

## 0.2.3 (2026-10-22)

NVIDIA key rotation. Upgrade with `pip install -U voltage-verify` before verifying a bundle
offline. Nothing that 0.2.2 reported as verified needs to be verified again.

* What changed at NVIDIA: since October 2026, NRAS also signs tokens with keys issued by a
  second intermediate, `CN=NVIDIA Attestation Service GPU Intermediate 005` (valid 2026-10-08
  to 2030-10-08). On 22 October 2026, NRAS's JWKS presented 77 of its 213 keys under 005.
  Both 004 and 005 verify under NVIDIA's root `NVIDIA Attestation Service CA 001`, as served
  by NVIDIA's PKI at
  `https://prod.vault.nvidia.com/v1/dgxc-attestation/pki/nras-prod/ca_chain` (checked on
  22 October 2026). The tool itself does not fetch or check that root.
* What 0.2.2 did with such a bundle: offline, `nvidia.signatures` failed, since only 004 was
  pinned and there is no live JWKS to vouch for another intermediate. Online it passed,
  through the intermediates NVIDIA's live JWKS presents. Failing closed, it never accepted a
  key it should not have.
* What 0.2.3 does: pins 005 next to 004 (SPKI SHA-256 `ba14d003...`, read from the live NRAS
  JWKS over HTTPS and cross-checked against confidential-dot-ai/attestation-rs). Everything
  else in the chain check is unchanged: a leaf issued by the intermediate, named NVIDIA
  Corporation, holding exactly the JWK's key, valid when the token was issued, listed once.
* The `nvidia.signatures` line now says "a pinned NRAS intermediate".

## 0.2.2 (2026-10-21)

Security release for the Intel TDX side of the verifier. Upgrade with
`pip install -U voltage-verify`. The NVIDIA checks were not affected: the GPU policy has always
required `dbgstat` disabled.

* What 0.2.1 and earlier accepted: a quote from a Trust Domain running in debug mode. Bit 0
  (DEBUG) of the TD attributes was listed in the `--json` facts (`td_attributes`) but never
  checked. A debug TD gives no confidentiality: the host can read and modify its memory and CPU
  state, and Intel's Quoting Enclave still signs its quote, so `quote` and `verify` reported
  it as verified. Found in an internal review on 21 October 2026.
* What 0.2.2 refuses: a new check, `tdx.debug`, in `quote`, `verify` and `selftest`, online
  and offline, fails when bit 0 of the TD attributes is set, and prints the TD attributes
  either way. It reads the TD report itself, so it is reported even when the signatures fail.
* `--allow-debug` (`quote` and `verify`) turns that failure into a warning, for development
  VMs only, the way `--accept-out-of-date` does for the TCB status.
* To re-read a result from 0.2.1 or earlier without running it again: if the first byte of
  `td_attributes` in the `--json` facts is odd (`01`, `03`, ...), the TD was in debug mode.
  The reference bundle in `examples/hello-workload/` and the test quotes read
  `0000001000000000` or `0000004000000000`: debug off.
* Library: `tdx.TdReport.debug`, `tdx.TD_ATTRIBUTES_OFFSET` (168) and the `allow_debug`
  argument of `verify_bundle` and `verify_quote`.

## 0.2.1 (2026-10-09)

Security release for the NVIDIA side of the verifier. Upgrade with
`pip install -U voltage-verify`. A bundle that passed with 0.2.0 or earlier proves nothing
about NVIDIA until it is verified again with 0.2.1. The Intel TDX checks were not affected.

* What 0.2.0 and earlier accepted:
  * With `--offline`, and online whenever NVIDIA's live JWKS could not be fetched, the JWK
    coordinates shipped in the bundle were trusted as they were; x5c was only read to display
    certificate subjects. A bundle could carry its own signing key, and NRAS tokens re-signed
    with it verified. This case was reported privately on 23 September 2026 by Satyawan Singh,
    who confirmed on 23 October 2026 that 0.2.3 refuses his original reproduction (bundle key
    without an x5c chain) and still verifies a genuine bundle. Thank you.
  * A JWKS listing the signing key id twice: the first entry was used.
  * Online, signing keys came from NVIDIA's live JWKS, not from the bundle: a bundle whose keys
    were missing or stripped of their chain still verified while NVIDIA published the key.
    The other way round, honest archived bundles failed online once NVIDIA had rotated the key
    (about two days later).
  * A token without `kid` was checked against the first key of the JWKS.
* What 0.2.1 refuses. These are the rules of the website verifier (voltagegpu.com,
  `lib/verified/nras.ts`):
  * Signing keys are taken from the bundle's JWKS only. A key that is only in NVIDIA's live
    JWKS is not used. Online, the live JWKS only contributes the intermediates it presents,
    so that a rotation of NVIDIA's intermediate does not break verification.
  * Each key needs an x5c of at least two certificates. The leaf must be issued by the pinned
    NRAS intermediate (SPKI SHA-256 `fd32837f...e896`, "NVIDIA Attestation Service GPU
    Intermediate 004") or by one that NVIDIA's live JWKS presents. It must name NVIDIA
    Corporation and hold exactly the key the JWK declares. Both certificates must have been
    valid at the token's `iat`.
  * Also refused: a key id listed more than once, a token without a usable `kid`, any
    algorithm other than ES384/ES256 (`none`, HMAC, RSA, any other name), and an algorithm
    that does not match the key's curve.
  * Malformed tokens, JWKS entries or x5c chains now fail the `nvidia.signatures` check
    instead of stopping the verifier with an exception. Tokens over 16 KiB are refused.
* Honest bundles are unaffected: the published evidence bundles still verify, offline and
  now also online.
* `attest` always embeds NRAS's JWKS, also with `--no-collateral`.
* `selftest` runs three more negative tests: `nras-own-key`, `jwks-duplicate-kid` and
  `jwks-no-chain`.

## 0.2.0 (2026-09-18)

* New command `voltage-verify quote <file>`: verifies a bare Intel TDX quote from any provider,
  raw, hex or base64, with no manifest and no NVIDIA token. Structure, signature chain up to
  the pinned Intel root, TCB status, QE identity and revocation from Intel PCS. Options
  `--report-data` (bind to the 64 bytes you expect), `--offline`, `--accept-out-of-date`,
  `--json`. Tested against Google's production Sapphire Rapids quote from go-tdx-guest, now
  vendored under `tests/fixtures/third-party/`.
* `tdx.parse_quote(raw, allow_trailing=True)` tolerates bytes after the signature data and
  reports how many were ignored (`Quote.trailing`); the default stays strict, so bundles
  produced by `attest` are checked exactly as before.
* A platform whose TCB components match no level in Intel's TCB info is reported as OutOfDate
  (Intel's own rule) and, with `--accept-out-of-date`, becomes a warning instead of a failure.

## 0.1.2 (2026-09-18)

* Docs only, no code change. Spells out what the Protected PCIe reference tokens in
  `examples/hello-workload/` actually cover (CC State OFF reading, one token per GPU, no NVSwitch
  fabric attestation). Adds a "Works with any Intel TDX host, any provider" section: `attest` and
  `verify` never call a VoltageGPU endpoint, they read the kernel's own configfs TSM interface and
  NVIDIA's own SDK and collateral, so the tool works the same on any TDX host with a confidential
  NVIDIA GPU, on any cloud.

## 0.1.1 (2026-09-12)

* Dependencies: the upper bound `cryptography<47` is removed. It existed only so that the
  `[attest]` extra could co-install with nv-attestation-sdk (which pins `cryptography==43.0.1`),
  but it also stopped verifier machines from getting current cryptography releases and their
  security fixes. The lower bound is unchanged; the test suite passes on 43.0.1 and 50.0.1.
  No code change.

## 0.1.0 (2026-09-12)

First release.

* `manifest`: image digest resolution through the OCI registry API (no pull), artifact
  hashing, free-form extras, verifier-side random challenge.
* `attest`: TDX quote through configfs TSM with `report_data = SHA-512(manifest)`, NVIDIA
  NRAS attestation with `nonce = SHA-256(manifest)`, single-GPU and multi-GPU Protected PCIe
  modes, Intel and NVIDIA collateral embedded for offline verification.
* `verify`: quote structure, attestation key signature, Quoting Enclave binding and signature,
  PCK chain to the pinned Intel SGX Root CA, TCB level and TDX module identity against Intel's
  signed TCB info, Quoting Enclave identity, CRLs, NRAS ES384 signatures, nonce, claims policy,
  token freshness, challenge and image expectations.
* `selftest`: six negative tests (image digest, artifact, manifest field, challenge replay,
  quote tampering, GPU claim tampering).
* Tests against real evidence captured on VoltageGPU Confidential VMs.
