I use AI agents to run my own business before I use them in anyone else's. If you're deciding whether to let me near your buying decisions, you should be able to see the rules I work under, what has been checked and what hasn't been yet.
Trust covers four things here: security, governance, compliance and best practice. Each has three columns:
A card moves up only when the evidence does. This page will grow.
How the agent environment is protected.
Written rules restrict sharing logins and keys between tools. They require each tool to use only its authorised sign-in or credentials. Technical enforcement across the estate hasn't been verified. That's a rule today, not a technical wall. The sealed-box evaluation below is the step towards a wall.
In one controlled test, a backup model finished the job after a simulated usage limit on the main model. The same job with no backup failed. That proves that path. It doesn't prove every kind of outage.
I've designed a synthetic-only evaluation of agent containment, using NVIDIA OpenShell, an open-source sandbox for AI agents. The idea: the agent works inside a box, and the limits on which files, websites and keys it can reach sit outside it. Instructions tell an agent what it should do. A box would limit what it can do. Nothing has run yet. When it does, this card will name the setup, what was blocked and what wasn't covered.
The UK government-backed baseline for cyber security. We're working towards it on a self-led route, with a baseline readiness checklist in place. This card changes when the certificate is issued.
Who decides, and how decisions are kept.
Agents prepare emails and posts. Nothing is sent or published externally unless I've reviewed it or set the authority for it in writing.
Founder decisions are recorded as dated rulings, with what I decided kept apart from what was inferred.
A routing-control test recorded 21 passing checks: when a reviewed source file was changed, missing or redirected, it was refused or held rather than used. That covers the setup tested, not every agent in every configuration.
Security and governance claims need specialist review before release. This page is going through it.
Each test gets an ID, a hypothesis, what would prove it wrong, the result and the evidence, including the failures. The format is drafted and the first past tests are written up. It isn't live yet.
Once the register is live, cards here will link to a plain-English summary of the test behind them.
What happens to your data.
Registered with the ICO, registration ZC141759, renewal due May 2027. Privacy questions go to privacy@techbridgeadvisory.co.uk.
Standing rules require client separation and the minimum data a task needs. Credentials must never be entered into AI tools.
I use an outside AI model to sort public research into topics. Admission checks pass eligible public-source records and exclude email, meeting transcripts, internal material and client material. Across 1,742 records, they matched the reference rule on every one. The scheduled daily run is a separate test still to be proven.
In progress. The AI use and data-handling policy is still a draft. I'll update this card as each piece of readiness work is signed off.
Every AI model and tool an agent can call, what it may receive and what it must never receive, in one record I can show you.
The standards I hold myself to, and only the ones I can show.
Written rules require minimum data and each tool's own sign-in. A key is only for the tool it was issued to.
External claims require evidence review. Vendor marketing doesn't count as evidence. If a number can't be supported, it doesn't ship.
The tests written up so far keep their failures next to their fixes. A green result that hides a red one isn't evidence.
The sealed-box evaluation is designed to try the escapes, not just the happy path: file traversal, blocked websites, direct addresses, leaked dummy keys and an agent trying to widen its own rules. If any of it gets through, it fails and stays failed until fixed.
Cyber Essentials first. An ISO 27001 readiness review once the platform's scope is settled. Neither is held today.
No security is absolute, and I won't present mine that way. I'll call something certified when the certificate exists. Until then each card says what the control does, how it was checked and where it stops.
