§ 01 · Mission
The verification gap
is a solvable problem.
Enterprise AI adoption is constrained by a verification gap. Existing approaches ask whether a response seems right — a weak test that fails precisely when the stakes are highest.
We solve this through surface-based detection: measuring the behavioural signature of model inference at the API boundary. Token statistics, timing patterns, and output characteristics carry reliable information about output quality. We extract that signal per-invocation, in real time, without ever accessing content.
The defensible claim
“We measure behavioural signals at the API boundary and detect deviations from calibrated baselines — without ever accessing prompt or response content.”
Every term is operationally defined. That's the point.
§ 02 · Team
The people behind Arkheia
David Murfet
Founder & CEO
CISO background with deep expertise in trust, security, and enterprise risk management. Led security transformations at scale. Patent inventor for AI integrity detection methodology. Security and BPR Consultant for 20 years. Former investment banking background with Lehman Brothers and Deutsche Bank. Enterprise experience spanning West Bromwich Building Society, COLT Telecom, Sompo International Insurance/Reinsurance, Conga, and Aptos.
Advisors & Team
Building
We're assembling a team of AI safety researchers, enterprise sales leaders, and security experts who share our mission to make AI trustworthy at scale.
§ 03 · Intellectual Property
Core patent portfolio
Patent 1
Behavioural Surface Profiling
Provisional Application Filed- ›Method for per-model baseline characterisation and deviation detection
- ›Validated: per-invocation risk scoring across heterogeneous model families
Patent 2
Surface Signal Detection
Validation Complete (Feb 2026)- ›Detection of abnormal outputs via API-boundary surface signals
- ›Validated across 4 provider families with strong signal separation
- ›Filed within 30-day post-validation window
§ 04 · Journey
From discovery to production
November 2025
Surface Signal Discovery
Identified that inference-time behavioural signals — token statistics, timing patterns, entropy — carry reliable information about output quality. The approach operates at the API boundary and requires no model access.
December 2025
Multi-Model Characterisation
Built characterisation corpora across Claude, GPT-4o, and major model families. Developed per-model behavioural profiles from empirical data — not generic thresholds.
January 2026
Extended Validation — Grok & Gemini
Validated detection profiles for Google Gemini and xAI Grok families. Strong signal separation confirmed across all providers. API proxy architecture designed and built.
February 2026
Production Proxy Deployed
MCP Trust Server launched — tool-native detection for Claude Code, Cursor, and agent frameworks. 100 tests, 7/7 pilot criteria passing. 13 model profiles across 4 providers, each with independent validation corpora.
March 2026
Enterprise Proxy & MCP Control Plane
Enterprise Proxy released — on-prem, air-gap capable deployment with hash-chain audit log, secrets redaction at boundary, and default-deny tool registry. MCP Control Plane adds tool-gating: detection signal becomes a hard control surface before execution, with run manifests and inherited integrity policy across all sub-agents.
§ 05 · Core Principles
What we stand for
01
Scientific Rigour
We treat AI as an engineered system with measurable observable behaviour. Detection profiles are built from controlled characterisation runs — not generic thresholds. Every claim is backed by validation data.
02
Privacy by Design
Detection operates entirely on behavioural signals at the API boundary. Prompts and responses are only transiently processed and never persisted. Privacy is structural, not a policy commitment.
03
Execution Speed
From characterisation corpus to validated detection profile in days. We move at the velocity of model deployment — not the velocity of enterprise software cycles.
Join us
We're building the trust layer for AI. If you're passionate about making AI safe for enterprise deployment, let's talk.
Get In Touch