Evidence-based
Every finding is reproducible and tied to a specific probe and response — not a subjective score.
What this actually is
A bid, not an award — told straight
The Sovereign AI R&D Procurement Scheme is a real government fund — £100m total, individual contracts typically £1–3m (range £250k–£10m) — but it is not a grant handed out on request. It is a competitive process: an Expression of Interest, then a scored Full Application against published criteria, run in bimonthly batches. We keep 100% of our IP either way, and payment is against delivery, not equity. This page is our case for why PrintzAI Red Team should win a place in it — we have not been awarded anything yet, and we will not pretend otherwise.
Why Challenge 3
“Enabling safe AI agent adoption” is the challenge we already live in
“Produce an evidence-based, operational, and ideally automated risk management approach for CISOs to support risk-based decision-making for specific agents in specific contexts. The risk management approach should include both resilience and cyber security risks. Alternatively, or in addition, develop an evidence-based approach to reducing operational risk of specific agents through architecture, monitoring, and controls, while supporting agentic adoption and innovation at pace. Solutions that support wide-scale adoption, such as those built on open-source frameworks, will be preferred.”
— Challenge 3, Sovereign AI R&D Procurement Scheme (NCSC-owned)
Read that back against what Red Team already does. Evidence-based — every finding traces to a specific probe and a hashed response, not a subjective score. Operational — it runs against a live agent on demand, not a one-off audit. Resilience and cyber-security risk — denial-of-wallet and tool-misuse sit next to prompt injection and data exfiltration in the same report. Architecture, monitoring and controls — that's a description of the kill switch, tool allowlisting and audit trail already running our own agent fleet. And it's built on a self-hosted, open-source model stack by default — the exact preference the challenge states. We are not proposing to build this from a standing start; we are proposing to fund what we have already shipped.
Operational
Built for continuous, repeatable use against a live agent fleet, not a one-off audit.
Open by preference
Built on a self-hosted, open-source model and tooling stack.
Already running
Not a proposal — a shipped feature, tested against our own production agents.
Proof, not promises
Real output from the live product
These are unedited screenshots from PrintzAI Red Team running against our own internal agent (codename “Breaker”) and an external endpoint, taken on 4 September 2026.
What £1,000,000 builds
A draft allocation — not yet final
This is a starting split, sized against what Red Team actually needs to go from a working prototype to an NCSC-grade submission. It will be refined before anything is submitted.
35%
£350,000
Security R&D & probe engineering
Two to three specialist security/ML engineers taking the probe library from today's 13 probes to comprehensive, continuously-updated adversarial coverage across prompt injection, jailbreak, tool misuse, data exfiltration and denial-of-wallet. This is the R&D core the scheme is funding.
25%
£250,000
Production-grade orchestration
Replace today's single-box job runner with a real distributed scan queue, multi-tenant scale, and a continuous monitoring mode — so an agent gets re-tested automatically every time its prompt or tools change, not just once.
15%
£150,000
Compute & open-model infrastructure
Dedicated GPU capacity so testing is not limited to one office PC, and so Red Team can test against the widest possible range of open-source models — directly answering the scheme’s stated preference for solutions built on open frameworks.
10%
£100,000
Independent validation & assurance
An external security audit of Red Team itself, and work toward the compliance evidence — Cyber Essentials Plus and equivalent — that a government supplier and enterprise customers will reasonably expect.
10%
£100,000
Pilot with the challenge-owner department
Dedicated engineering time to run Red Team against a real public-sector AI deployment under NCSC guidance, and fold what we learn straight back into the product.
5%
£50,000
Commercialisation for UK business
Documentation, pricing and go-to-market so every UK company deploying an AI agent — not only the ones bidding for government contracts — can put Red Team in front of it.
The road ahead
Milestones
NOW
PrintzAI Red Team is live
Shipped to production and already run against PrintzAI’s own agent fleet — the evidence on this page is real output, not a mockup.
NEXT
Expression of Interest
Submitting for Challenge 3 — "Enabling safe AI agent adoption" (NCSC) — under the £100m Sovereign AI R&D Procurement Scheme. First application batch closes 1 October 2026.
THEN
Full Application
If shortlisted from the Approved Supplier List: a scored submission on technical approach, delivery plan, impact and value for money.
IF FUNDED
The £1M roadmap, plus a pilot
Twelve months delivering the breakdown below, with a pilot deployment against a real public-sector AI system.
BEYOND
Open it up
Take the resulting risk-management framework to every UK business deploying AI agents — not just the public sector.
About me
Cornelius Wilkinson — built in Welton, East Yorkshire, built for family
I'm Cornelius, founder of Fresh Printz Automations and PrintzAI. I studied electrical and electronics engineering before eight years in the military, then came home to be with my son Oliver, my wife Melissa, our daughter Phoebe, and our beagle Mable. From there into custom engineering work — and the same realisation each time: nobody was building a way out of the grind, so I built one myself.
That background is why Red Team is not a pitch-deck idea. Eight years of military discipline is why the product ships with a kill switch, an audit trail and an authorization gate before it will touch anything it doesn't own — engineering restraint isn't a slide, it's a habit. The engineering background is why I personally built and run the self-hosted model infrastructure this whole platform stands on, GPU box and all, rather than outsourcing it. And the family is why this matters to me beyond the contract: I want PrintzAI to give people time back, and I want to build something in the UK that a national cyber-security agency can trust, not just a startup that talks a good game.