PRIVACY

§ Arkheia · Security & Privacy

Security &
Privacy

We measure signals, not content. Privacy isn't a policy — it's structural. Synesis runs inside your own infrastructure. Your data never leaves your perimeter.

§ 01 · Two products, two data models

Detection proxy and governance platform
handle data differently by design

Detection Proxy

Signals only — no content

The API proxy sits inline in your model calls. It extracts behavioural signals — token probabilities, timing, entropy — during transit. Nothing is stored. The signals are computed and discarded. Your prompts and responses pass through unchanged.

Data processed: behavioural telemetry only · retained: zero · content access: none

Synesis Governance Platform

Self-hosted — your infrastructure, your data

Synesis runs entirely inside your own infrastructure. The database, the audit trail, agent registry, signals, compliance evidence — all of it lives in your environment. Arkheia has no access to your governance data. Ever.

Deployment: docker-compose or Kubernetes · data residency: your perimeter · cloud telemetry: none

§ 02 · Detection — Core Principle

We measure signals.
Not content.

Detection operates at the signal level — token statistics, timing patterns, entropy. These are observable at the API boundary without accessing what was asked or what was answered.

This isn't a commitment we made after designing the system. It's the architecture. The detection method operates on behavioural signals — it doesn't require access to your prompts or responses.

The analogy

“Monitoring CPU temperature doesn't reveal what code is running. Arkheia measures behavioural signals without ever seeing the conversation.”

We monitor the engine, not the conversation.

What we observe

  • Token generation timing and inter-token latency
  • Token log-probability distributions (from provider API)
  • Response entropy and output statistics
  • Invocation metadata (timing, token counts)
  • Hardware execution signatures (on-prem deployments only)

Extracted during transit. Never stored — discarded after signal computation.

What we never observe

  • User queries or prompt content
  • AI-generated response content
  • Personal identifiable information
  • Business logic or proprietary data
  • Model weights or training data
  • Application context or session history

§ 03 · Synesis — Governance Data Model

Your governance estate. Your data.

Synesis is self-hosted. The full governance dataset — agents, policies, signals, compliance evidence, kill switch journal, audit trail — lives in your own PostgreSQL instance. Arkheia has no access to any of it.

Agent Registry & Policies

Agent identities, autonomy tiers, authorisation trails, and policy definitions are stored in your database. You own the schema. You control access via RBAC (demo / operations / cxo / admin).

Signal Feed & Kill Journal

Every signal — UNAUTHORIZED_MODEL, PROMPT_INJECTION_DETECTED, GUARDRAIL_BYPASS, PII_DETECTED — is written to your event store. The kill switch journal is immutable and stays in your perimeter.

Compliance Evidence

Compliance reports (EU AI Act, NIST AI RMF, SOC 2, GDPR, ISO 42001, OWASP Agentic) are generated from your own governance data and signed with your own timestamps. No third-party dependency.

Access Control

demo

Read-only access to all sections. Scoped kill switch (model/workflow/agent/MCP).

operations

Signal acknowledgement, agent recertification, scoped kill switch.

cxo

Tenant-level kill switch. Full posture visibility.

admin

Full platform access including tenant-level halt.

Compliance Coverage (Live Demo)

EU AI Act
71%
NIST AI RMF
75%
SOC 2
80%
ISO 42001
60%
GDPR
63%
OWASP Agentic
58%

Coverage generated from your own workflow data. Increases as you act on improvement opportunities.

§ 04 · Deployment Options

You choose your security posture

API Proxy

Detection
  • Signals extracted at the proxy layer — no content access
  • Zero prompts or responses stored
  • Encrypted transit (TLS 1.3)
  • Behavioural signals discarded after risk computation

Ideal for: Developer teams, agent pipelines, cloud-first deployments

Synesis — Self-Hosted

Governance
  • Runs inside your own infrastructure (docker-compose or Kubernetes)
  • All governance data stays within your perimeter
  • You own the database, the audit trail, and the keys
  • No data sent to Arkheia — zero telemetry to cloud

Ideal for: Enterprise governance, regulated industries, sovereignty requirements

On-Premise Engine

Air-gapped
  • All data stays within your infrastructure
  • Zero signals leave your network
  • No external API calls from the engine
  • Complete audit trail under your control

Ideal for: Regulated industries, sovereign requirements, high-security environments

§ 05 · Compliance & Certifications

Regulatory posture

Synesis maps every governance control to the regulatory article it satisfies. Coverage is live — it reflects your actual workflow data, not a static claim. All six frameworks are tracked continuously.

EU AI Act

71%
  • Art. 9 Risk Management — COMPLIANT (signal engine + posture score)
  • Art. 12 Record-keeping — COMPLIANT (immutable audit trail, TimescaleDB)
  • Art. 13 Transparency — PARTIAL (model registry; end-user notice pending)
  • Art. 14 Human Oversight — COMPLIANT (kill switch + signal acknowledgement)
  • Art. 17 Quality Management — PARTIAL (incident playbooks; review cadence pending)

NIST AI RMF

75%
  • GOVERN 1.4 — COMPLIANT (RBAC, JWT, tenant isolation, audit log)
  • GOVERN 4.1 — COMPLIANT (model authorisation SoD, kill switch, OPA gateway)
  • MAP 1.3 — COMPLIANT (trust zone topology, signal categories)
  • MAP 1.5 — COMPLIANT (agent certification, autonomy tiers, signal rules)
  • MEASURE 2.1 — COMPLIANT (posture score 0–100, 7 and 30-day trend)
  • MANAGE 6.1 — PARTIAL (incident playbooks + kill switch; drill records pending)

ISO 42001

60%
  • A.9 Risk Controls — COMPLIANT (OPA gateway, kill switch, SoD enforcement)
  • A.10 AI System Documentation — COMPLIANT (model registry, agent registry, RAG sources)
  • A.6.2 Impact Assessment — PARTIAL (autonomy tiers + golden images; DPIA form pending)
  • C.7 Competency — PARTIAL (RBAC implies training; formal records not tracked)

SOC 2

80%
  • CC6.1 Logical access controls — COMPLIANT (JWT HS256, section RBAC, tenant isolation)
  • CC6.3 Role-based access — COMPLIANT (6 roles, section matrix, SoD enforced)
  • CC7.3 Incident response — COMPLIANT (signal status workflow, kill switch, operations section)
  • CC8.1 Change management — COMPLIANT (model/MCP changes require SoD approval)
  • CC4.1 Anomaly monitoring — COMPLIANT (signal engine, 41 signal types, posture score)

GDPR

63%
  • Art. 32 Security of processing — COMPLIANT (JWT, tenant isolation, append-only audit log)
  • Art. 25 Data protection by design — PARTIAL (PII_DETECTED signals; redaction not automated)
  • Art. 30 Records of processing — PARTIAL (invocation log exists; RoPA format not generated)
  • Art. 35 DPIA — PARTIAL (agent risk tiers + trust zones; formal DPIA workflow pending)
  • Detection: behavioural signals are operational metadata — not personal data under GDPR

OWASP Agentic AI

58%
  • OA-01 Agent identity & authorisation — COMPLIANT (agent registry, cert status, RBAC)
  • OA-02 Tool scope enforcement — COMPLIANT (MCP registry, SoD approval, prohibited tools)
  • OA-03 Human oversight — PARTIAL (kill switch + signal ack; structured oversight record pending)
  • OA-04 Supply chain integrity — PARTIAL (golden images, drift detection; signing pending)
  • OA-05 Prompt injection detection — PARTIAL (proxy fires signal; native prevention not in Synesis)
  • OA-06 Agentic audit trail — GAP (single-agent logged; multi-agent chain tracing not yet built)

Certifications

SOC 2 Type II in progress · ISO 27001 planned · Third-party penetration testing conducted

Coverage model

Weighted: COMPLIANT = 1.0 · PARTIAL = 0.5 · GAP = 0. Increases as you act on improvement opportunities in Synesis.

Self-hosted advantage

All compliance evidence generated from your own data. No third-party dependency. You sign the reports.

§ 06 · Stakeholder Messaging

How to brief your stakeholders

For CISOs

We're instrumentation, not surveillance. The detection proxy processes behavioural signals — the same category of data your network monitoring tools collect — without accessing conversation content. Synesis runs inside your own infrastructure.

For Legal & Compliance

Behavioural signals are not personal data under GDPR — they're operational metadata equivalent to server performance logs. Synesis governance data (signals, policies, compliance evidence) stays in your own database. You remain the data controller for everything.

For Procurement

Synesis is self-hosted: git clone, docker-compose up, running in under two minutes on your own infrastructure. No data leaves your perimeter. The detection proxy offers API, on-prem, and isolated deployment options. SOC 2 Type II in progress.

The challenge

Can you reconstruct an AI query from token timing and probability patterns?

No.

We measure signals. Synesis keeps your governance data in your own infrastructure. You keep your data.

§ 07 · Audit & Traceability

Operational evidence remains auditable

Every invocation through the Arkheia detection proxy produces structured artifacts. In Synesis, every governance action — signal acknowledgement, agent recertification, kill switch fire, report generation — is recorded immutably in your own audit trail.

Detection Proxy Artifacts

  • run_id
  • Model identifier + version
  • Behavioural surface version
  • Zone decision
  • Remediation actions taken
  • Timestamps

Synesis Audit Trail

  • Signal acknowledgements (who, when, action)
  • Agent certification changes (CERTIFIED / UNCERTIFIED)
  • Kill switch fire events (scope, reason, actor) — immutable
  • Compliance report generation (framework, date range, signatory)
  • Policy changes and tool approvals
  • RBAC role assignments

Detection artifacts can be exported to SIEM, ticketing, and compliance systems.

Detection internals remain private. Governance evidence stays in your own infrastructure. Everything is auditable.

§ For Your Security Team

Security documentation available

For detailed security architecture, penetration test results, compliance certifications, or a technical briefing for your security team — contact us directly.