# Banbreach

> Security engineering and product discipline, applied to AI agent workflows.

Banbreach is a boutique consultancy that helps organisations turn promising AI agent workflows into controlled, testable and economically justified business operations. The principal does the work personally — backgrounds in cyber security engineering and product development, not a delivery bench.

Every claim on the site is measured, sourced, or explicitly labelled as an estimate or hypothesis. Articles state what their method can and cannot establish. Content on this site is general method description, not advice for a specific system.

## Services

- [AI Workflow Assessment](https://www.banbreach.com/services/workflow-assessment/): One consequential workflow, ten business days — safety, security, efficiency and economics examined, with a threat model, control analysis, acceptance criteria and release gates as deliverables.
- [Executive Workshop](https://www.banbreach.com/services/governing-ai-agents/): A three-hour working session giving leadership teams a practical model for what "under control" means for AI agents.
- [AI Incident Response and Forensics (DFIR for AI)](https://www.banbreach.com/services/ai-incident-response/): Scoped engagements that reconstruct incidents involving AI agent workflows — what the agent did, what it was permitted to do, which controls failed — plus an Incident Readiness Review that makes workflows investigable before an incident. Urgent route details and honest limits are stated on the page.

## Pages

- [Home](https://www.banbreach.com/): Positioning, workflow readiness check (runs entirely in the browser), and entry points.
- [Methods & Demonstrations](https://www.banbreach.com/methods/): Ten methods presented uniformly — problem, how it works, what it can establish, what it cannot, how it fails, and a worked example. Threat modelling and release criteria lead; forensic readiness covers incident analysis.
- [Insights](https://www.banbreach.com/insights/): Articles by the principal on AI security, workflow assessment, agentic design patterns and decision economics.
- [About](https://www.banbreach.com/about/): The firm, the principal, the verification policy for published claims, and where the approach comes from.
- [Press & media](https://www.banbreach.com/press/): Verifiable media mentions only, plus the speaking and media kit — bios, headshots, topics and the press contact route.
- [Contact](https://www.banbreach.com/contact/): Enquiry routes; a principal replies within one business day on the planned routes; live incidents use the monitored urgent channel stated at the top of the page.

## Insights articles


## Policies

- [Privacy notice](https://www.banbreach.com/privacy/): What this site collects and why — no marketing email, analytics is consent-gated.
- [Cookies & analytics](https://www.banbreach.com/cookies/): The complete cookie inventory and consent controls.
- [Terms of use](https://www.banbreach.com/terms-of-use/): Enquiries are not engagements; content is informational, not advice.
- [Insights corrections policy](https://www.banbreach.com/insights/corrections/): How substantive errors in published articles are reported, verified and marked.
- [Accessibility](https://www.banbreach.com/accessibility/): WCAG 2.2 AA target, implementations and known limitations.
- [security.txt](https://www.banbreach.com/.well-known/security.txt): Vulnerability disclosure contact and policy.

## Feeds

- [RSS](https://www.banbreach.com/rss.xml)
- [JSON Feed](https://www.banbreach.com/feed.json)

## Synthetic sample artefacts

- [Sample assessment dossier](https://www.banbreach.com/sample-dossier/) and [sample workshop pre-read](https://www.banbreach.com/sample-pre-read/) are fully synthetic documents. They describe no real client, workflow or result, and must not be cited as evidence of outcomes.

## Notes

- Dates: articles carry publication and update dates; substantive pages carry a last-reviewed date.
- Corrections to factual or technical claims in articles are visibly marked, per the corrections policy — do not silently restate corrected claims.
- Site copy uses British English spelling (e.g. "organisation", "modelling").