Skip to content
Kernwall
Trust · for your security team

Don't trust us. Check every release offline.

Your security team verifies every Kernwall OS release offline. Every request leaves a record, and nothing calls out. Below are the threat model, the supply chain and a dated plan for each certificate.

Threat model

Five ways in, and what stands in each.

Kernwall OS draws the trust boundary at your site wall and treats everything that crosses it as hostile, us included. Each threat names its control and the record your assessor gets.

YOUR SITE · TRUST BOUNDARYIMPORT STATIONTPM 2.0 · SECURE BOOT01Hardware root of trust02Immutable OS03Model runtime2026.10 · VERIFIED04Cross-gap deliveryKERNWALL05Accreditation engineAIR GAP · NOTHING ROUTES OUTINTERNETNO ROUTEBUNDLE · SIGNEDImport stationACCEPTED · RE-SIGNED FOR THIS SITE1234523
Trust boundary at your site wall · threats pinned where they enter
  1. Insider

    Attack
    An admin or user reads data they shouldn't, or quietly changes a model or a log.
    Control
    The OS image is read-only, so changing it takes a signed bundle, not a login. Retrieval inherits source-system ACLs. Every prompt, retrieval and tool call joins a hash-linked chain tied to user identity.
    Evidence
    The audit chain, anchored to TPM quotes and exported by media or diode.
  2. Supply chain

    Attack
    A tampered OS image, driver, container or model arrives on the update media.
    Control
    One signed release bundle. The import station verifies signatures against an offline trust root, re-scans, and re-signs with your site key before anything reaches the appliance.
    Evidence
    The verify report, SBOM and scan report for each bundle imported.
    See the update flow
  3. Malicious model

    Attack
    Weights that execute code on load, or a fine-tune of a base your rules bar.
    Control
    Ingest accepts safetensors or GGUF only, with no pickle and no unreviewed remote code. Lineage is checked against policy, and every promotion passes eval and red-team gates.
    Evidence
    A signed AI-BOM and eval report for each model you serve.
    See the lineage check
  4. Prompt injection

    Attack
    A document in the RAG index tells the model to leak data or call a tool.
    Control
    Retrieval is limited to what the asking user may read, chunks carry classification labels, and tools run only from signed, allow-listed manifests. Egress is denied, so there is nowhere to send data to.
    Evidence
    Tool calls and retrieved document IDs in the audit chain. Mitigated, not solved: no vendor can promise more.
  5. Physical access

    Attack
    Someone pulls a disk, boots their own kernel or swaps the root filesystem.
    Control
    Owner-enrolled Secure Boot keys, a signed unified kernel image, a dm-verity root, and a data volume sealed to the TPM's signed boot policy. Lockdown and signed-only modules are on.
    Evidence
    Measured-boot PCR values and a continuous TPM + IMA attestation record.
    Try to tamper with the boot chain
Out of scope, said plainly

Whoever holds your owner keys can sign what they like. Kernwall records every signing in the audit chain; it can't make one impossible. Implants in the server hardware itself belong to your procurement chain and the OEM. A model that is simply wrong is out of scope too: evals make it rarer, and your process still decides what a model may do.

Supply chain

Every byte on your appliance has a receipt.

Each release crosses the gap as one bundle: OS image, GPU drivers, runtime and model shards. Every step leaves a record your team checks offline, without trusting us.

  1. Source

    Pinned upstream sources, reviewed and mirrored. Nothing fetched at build time.

  2. Reproducible build

    Two independent builders produce the same bytes, or no release is cut.

  3. SLSA provenance

    A signed attestation of what was built, from which commit, on which builder.

  4. Signature

    TUF metadata, ML-DSA and LMS signatures, a trusted-time floor for expiry checks.

  5. SBOM · AI-BOM · VEX

    Every package and model listed, to the G7/CISA minimum elements for AI SBOMs; VEX says which CVEs actually reach the image.

  6. Site import

    Kernwall verifies it offline, re-scans it and re-signs it with your key. Your key decides what runs.

Checking a release offline, then signing it in
kernwall verify --offline · admin@import-01
$ kernwall verify --offline --root /media/usb/root.json /media/usb/release-2027.01.oci
tufroot v3 → targets v41 · expires 2027-07-01 ✓
time floor2027-01-14 from bundle · site clock later ✓
signatureML-DSA-87 + LMS · release key ✓
provenanceSLSA v1 · builder-a + builder-b agree ✓
digestos, drivers, runtime, 3 model shards match ✓
sbom1,284 packages · CycloneDX 1.6
vex17 CVEs present · 15 not affected · 2 fixed here
kev0 KEV-listed CVEs reachable ✓
$ kernwall import --sign-with site-key release-2027.01.oci
re-signedsite key · audit record #000412 ✓
staged to slot B · applies on next maintenance window

Certification roadmap

Every certificate we plan, in order.

Kernwall holds no certificate today. This is the order we pursue them in, and the page moves as each one starts.

Status as of 29 Oct 2026 · nothing held

  1. From the first release

    • CRA vulnerability handling

      Before first release · planned

      PSIRT, coordinated disclosure, security.txt and the ENISA reporting process. CRA reporting has applied to products on the EU market since 11 Sep 2026.

    • SBOM and VEX per release

      Before first release · planned

      CycloneDX or SPDX for every bundle, with VEX statements and the scan report beside them.

    • Secure development lifecycle

      planned

      Built to NIST SSDF and IEC 62443-4-1 from the start, so the later audit checks a process that already runs.

    • OSCAL and SCAP content

      planned

      Kernwall ships an OSCAL component definition and OpenSCAP checks for your assessor.

  2. First certificates

    • ISO/IEC 27001

      planned

      Information security management for the company that builds and signs the releases.

    • IEC 62443-4-1

      planned

      Secure product development, audited. The certificate your fab, utility or OT assessor asks for.

    • SOC 2 Type II

      planned

      The controls report your bank or hospital auditor asks for.

    • ISO/IEC 42001

      planned

      AI management system.

    • DISA Vendor STIG

      planned

      Intent form, then a vendor-written STIG for US assessors. Supports a faster ATO; it is not one.

  3. European government

    • ANSSI CSPN

      planned

      France's time-boxed first-level certification of the security functions, recognised by Germany's BSZ.

    • CRA conformity and CE marking

      planned

      Operating systems are CRA important products (Annex III, Class I). Required before the CRA fully applies.

    • EUCC (substantial)

      planned

      The only EU Common Criteria route since national schemes ended on 27 Feb 2026. Evaluation starts in this phase.

    • BSI VS-NfD approval

      planned

      Germany's restricted-data approval, through the qualified-manufacturer route.

    • IEC 62443-4-2

      planned

      Component requirements for the appliance as a host device, at SL2: the level ISASecure recommends as the minimum for off-the-shelf components.

    • NATO NIAPC listing

      planned

      Through a sponsoring nation, on the back of EUCC.

  4. Later, with a US partner

    • NIAP GPOS evaluation

      planned

      US Common Criteria for general-purpose operating systems, after the European certificates.

Faster ATO, never “ATO included”

Your assessor grants the authority to operate, and no vendor can ship one. Kernwall ships the evidence that shortens it: SBOMs, scan results, OSCAL documents and an audit chain mapped to the controls your assessor tests.

Documents and policies

What every release must carry.

Nothing here downloads yet, because no release exists yet. Kernwall calls a build a release only when it carries every item below.

Published with each release
  • Architecture and security whitepaper
  • SBOM, AI-BOM and VEX for every release
  • Scan reports for every release, beside a stock Ubuntu + CUDA image
  • Penetration-test summary from an outside lab
  • STIG and SCAP results
  • OSCAL component definition
  • Signed release feed and changelog
Policies from day one
Disclosure
A coordinated disclosure policy, security.txt and a PGP key, published when the security contact goes live. Until then there is no product to report against.
KEV fixes
Design target: a signed fix bundle on the day a CVE in the image joins CISA's KEV list, as its own emergency channel.
Telemetry
None by default. If you want health data out, it leaves through a one-way diode you control.
Vendor access
No remote access path for Kernwall. Support happens on site or through media you inspect.
Keys
Release keys sign bundles; your site key decides what runs. We never hold your site key.

Test it

Bring your assessor's questions to a pilot.

Your pilot runs on your hardware, inside your walls. Your security team checks every claim on this page.