MICKAI®Security
Team
The Sentinel, the armoured guardian of the Mickai security perimeter, before a fortress gate laced with gold circuitry
Security architecture

Security architecture. Control, records, verification.

8 control domains and 44 specific controls described for the Mickai Sovereign Intelligence Operating System. The design brings together operator-held keys, hardware attestation, agent controls and the Open Audit Record on customer-controlled infrastructure. Evaluate the supported configuration, egress boundary and audit coverage for your workflows. Signed records support integrity checks; they do not establish correctness, complete event capture or immunity from attack. No third-party certification is claimed here.

8
Control domains
44
Controls described
Local
Scoped processing
PQ
Signed audit (ML-DSA-65)
104
Filed UK patent applications
The subsystems behind the controls
Inspect the audit evidence
Open the offline audit verifier

Drag a Mickai Open Audit Record chain into the verifier. Hash linkage and FIPS 204 ML-DSA-65 signatures are checked in your browser. Load the verifier and required key material before testing disconnected use, and establish trust in the key separately. A valid record does not by itself prove completeness, authorisation or correctness.

→
The control domains
Define and test the processing boundary

Sovereign architecture and air-gapped deployment

Local inference, identity, governance and audit support isolated workflows when their dependencies are available inside the boundary. Assess telemetry, licensing, updates, backups and support tooling as well as the model. Approved connected work needs a separately scoped egress path and cannot be described as fully air-gapped.

  • Defined data-egress boundary
  • Disconnected workflow operation
  • Operator-commissioned egress only
5 controlsInspect the sovereign architecture and air-gapped deployment controls →
Inspect records, signatures and event coverage

Open Audit Record, tamper-evident post-quantum sealing

The Open Audit Record combines hash-linked records with operator-controlled ML-DSA-65 signatures. Verification can detect changes to recorded content subject to trusted keys and the checks performed. It does not establish that an answer was correct, every event was captured or an action was authorised; request evidence for each of those separately.

  • Per-action post-quantum signing
  • Hash-chained decision lineage
  • Browser-verifiable offline replay
5 controlsInspect the open audit record, tamper-evident post-quantum sealing controls →
Identity rooted in operator-controlled hardware

Hardware attestation, TPM 2.0 and measured boot

TPM 2.0, secure enclave and HSM configurations support hardware-rooted identity. Review which keys are hardware-protected, how devices enrol, and how measured boot is checked against an approved baseline. Attestation is evidence about measured state, not proof that the running system cannot be compromised.

  • Hardware-bound operator identity
  • Measured boot platform integrity
  • Hardware-bound licensing
6 controlsInspect the hardware attestation, tpm 2.0 and measured boot controls →
Inspect untrusted inputs and constrain actions

Agent safety and prompt-injection defence, Sentinel

Sentinel's design combines prompt-injection inspection, destructive-action checks, per-tool limits and approval gates. Test malicious document instructions, denied tools and approval failures against the proposed version. Inspection and isolation can reduce risk but do not establish that all injections will be detected or all harmful actions prevented.

  • Destructive-action interception
  • Prompt-injection inspection
  • Voice-biometric quorum on high-stakes actions
7 controlsInspect the agent safety and prompt-injection defence, sentinel controls →
Scope connected work and its permissions

Perimeter, sandbox and egress gateway

Connected workloads such as research or browsing are scoped to a sandbox and allowlisted egress gateway, with an operator-selected VPN where appropriate. Review process, file, credential and network permissions, including tenant boundaries. Test actual destinations and audit coverage rather than assuming every possible path is gated.

  • Sandboxed connected workloads
  • Allowlisted egress gateway
  • Per-tenant egress firewall
5 controlsInspect the perimeter, sandbox and egress gateway controls →
Make key custody and administration explicit

Key and identity sovereignty, operator-held keys and encryption at rest

The design places signing keys and the audit ledger under operator control. Confirm how keys are stored, backed up, recovered and revoked, and which administrators or support staff can access the deployment. Local custody does not itself remove legal obligations, insider risk or dependencies in the supply chain.

  • Operator-held signing keys
  • Vendor administration controls
  • Encryption at rest under operator keys
5 controlsInspect the key and identity sovereignty, operator-held keys and encryption at rest controls →
Test separation between people and departments

Multi-tenant isolation and access control

The access-control design includes tenant separation, partitioned records, row and column permissions, and clearance-gated skills. Confirm supported identity and re-authentication methods, including any voice-biometric controls. Test cross-tenant retrieval, cached content, account revocation and unattended sessions before relying on separation.

  • Cryptographic tenant isolation
  • Row and column access control
  • Clearance-gated skill execution
5 controlsInspect the multi-tenant isolation and access control controls →
Review provenance, versions and trusted signatures

Model and supply-chain integrity, signed weights and licence provenance

Signed artefacts, version pinning and licence provenance support controlled software and model updates. Identify the artefacts covered, trusted signing keys, dependency inventory and approval process. Test refusal of unsigned or altered packages, rollback and interrupted updates; a valid signature alone does not establish that software is safe.

  • Signed artefacts fail closed
  • Third-party-notice and SBOM hygiene
  • Signed model load lineage
6 controlsInspect the model and supply-chain integrity, signed weights and licence provenance controls →
Questions
What is the Mickai security model?

Mickai's SIOS security design combines customer-controlled infrastructure, operator-held keys, hardware attestation, agent controls and the signed Open Audit Record. An isolated workflow requires local models and supporting services; approved external connectors have a separate boundary. Confirm supported controls and event coverage for the proposed version, then test permissions, egress, failures and recovery. Architectural descriptions are not a security certification or proof of deployment behaviour.

How does Mickai defend against prompt injection and unsafe agent actions?

Sentinel is designed to inspect untrusted inputs and gate destructive actions. MickaiClaw provides per-skill authority controls, with dry-run and human approval mechanisms to assess for high-impact actions. Confirm which controls, including any voice-biometric quorum, are supported and configured. Test malicious context, bypass attempts, revoked authority and unavailable policy services. No detector, firewall or sandbox alone proves that prompt injection or data disclosure is impossible.

Who holds the encryption keys and the audit trail?

The design places key custody and the Open Audit Record with the operator. Request the actual key-management design: signing and encryption keys, hardware protection, administrators, support access, backup, recovery, revocation and any escrow arrangements. Independent audit verification needs the records, a trusted public key and a compatible verifier. Confirm these arrangements in the proposed deployment rather than inferring them from hardware ownership.

What can an independent reader verify in a Mickai audit record?

The browser verifier checks hash linkage and ML-DSA-65 signatures on supplied records. Load the verifier and required key material before testing disconnected verification, and establish trust in the key separately. A successful check supports record integrity and association with a signing key. It does not prove that all events were recorded, that an answer was correct or that an action was appropriate or completed as intended.

Shared responsibility

What to verify in the system, and what to operate.

Customer-controlled infrastructure still needs a defined responsibility model. Agree supplier and operator duties, validate the configured controls and retain evidence of how the estate is operated. Local hosting does not establish compliance.

Controls to verify in the deployment

  • Defined egress boundary and disconnected workflow operation
  • ML-DSA-65 signatures and the Open Audit Record's event coverage
  • Hardware-bound identity, measured boot and clone refusal
  • Destructive-action controls and prompt-injection defences, including failure cases
  • Signed artefacts that fail closed, and offline audit verification

Operated by your organisation

  • Physical security of the machines and the rooms they sit in
  • Custody and backup of operator keys, and operator vetting
  • Applying signed updates on your own schedule
  • Network segmentation around any commissioned egress
  • Incident response ownership, with the sealed audit record as evidence
Coordinated vulnerability disclosure

Found something? Tell us directly.

Mickai welcomes good-faith security research. Report suspected vulnerabilities to micky@mickai.co.uk with enough detail to reproduce the issue. We acknowledge reports within two working days, keep you informed while we investigate, and credit researchers who wish to be named once a fix ships. Please act in good faith: no data exfiltration, no service disruption, and no disclosure before we have had a reasonable window to remediate. A machine-readable policy is published at /.well-known/security.txt (RFC 9116).

For investors and evaluators

Inspect the controls and the evidence.

Request a briefing scoped to your workflows and the proposed version. Ask to see permitted and denied actions, audit verification, egress tests and recovery, with limitations and outstanding assurance work stated.

See the investor briefing