Your models, agents, and data are now a perimeter — and the things probing it are other models: tireless, adaptive, and cheap to run. A firewall from 2015 will not save you. A fort might.
DataFort is an AI safety and counter-AI attack practice. We audit your AI the way an attacker’s AI would, guard it at runtime, and — when the probing turns into an assault — we answer machines with machines: decoys, tarpits, and honey models inside your own walls, every countermeasure under human command.
An illustration of the idea, not telemetry: automated probes arrive from the left; the wall inspects and turns them away before they reach the keep. Rampart, below, is the real version of this picture.
The siege catalogue
These are not hypothetical. Each one is a documented pattern of attack against production AI — and each one has a counter. This is the map we work from.
Instructions smuggled into web pages, documents, emails, or tool output hijack your model mid-task — it obeys the attacker’s text instead of yours.
Injection test batteries against every input channel, plus runtime separation of trusted instructions from untrusted content.
An autonomous agent with tools — email, payments, code, databases — is steered step by step into doing the attacker’s work with your credentials.
The Agent Leash: tool allowlists, spend and step limits, and human gates on irreversible actions, enforced outside the model.
Patient, ordinary-looking questions walk your training data, system prompts, customer records, or RAG corpus out through the chat window.
Egress watch on outputs, canary records planted in your corpus, and extraction-pattern detection across sessions — not just single prompts.
A competitor or scraper queries your API systematically and trains a copy of your model — your R&D, rebuilt from your own answers.
Query-pattern throttling, output watermarking, and honey models: decoy endpoints that feed the cloner subtly wrong answers.
Cloned voices and faces of your executives, staff, or customers authorize payments, reset credentials, and talk their way past your people.
Provenance checks and out-of-band challenge protocols for high-stakes requests — a verification drill your team actually rehearses.
Not one attacker but a fleet: agents that probe, learn from each refusal, rephrase, and return — a siege that never sleeps and never gets bored.
Counterbattery operations: tarpits that waste swarm compute, decoy surfaces that absorb the assault, and defenses that adapt between waves.
The garrison
Engage one, or stack all three. Every engagement starts with the audit — you cannot defend a perimeter you have not walked.
A fixed-scope AI safety assessment run the way an attacker’s model would run it: automated probing at scale, then human analysts on the findings.
Fixed scope · typically 3–5 weeks · report + retest
Runtime defense that sits on the wall while your AI works: watching inputs and outputs, enforcing the Agent Leash, and raising the alarm in evidence-grade detail.
Standing service · monthly retainer · 24/7 watch floor
When defense alone is losing, we make attacking you expensive. Counter-AI operations run inside your perimeter only — deception, delay, and disruption aimed at the attacker’s machines.
Activated per campaign · human commander required · rules of engagement in writing
Doctrine
Counter-AI work lives one bad decision away from becoming the thing it fights. These four rules are written into every engagement, before any tooling is discussed.
Every countermeasure operates inside systems you own or control. We do not strike back at outside infrastructure, pursue attackers onto third-party systems, or “hack back.” Deception is lawful inside your own fort; it stops at your gate.
Machines probe, flag, and hold. A named human commander authorizes every active countermeasure, can halt one mid-flight, and answers for it afterward. No autonomous retaliation, no unattended escalation.
If it isn’t logged, it didn’t happen. Every probe, block, and countermeasure is recorded to a standard that survives scrutiny — from your board, your insurer, or a courtroom.
We defend whatever you run — frontier APIs, open weights, fine-tunes, agents built in-house. DataFort sells no model and takes no side in anyone else’s stack, so our findings never bend toward a product.
Defense builder
Tell us what you run and what you fear. The builder assembles a starting posture and drafts your briefing request — copy it or download it, and bring it to a founding-cohort conversation.
Pick at least one asset and one threat, and the wall takes shape.
Your stack will appear here — audit first, then the layers your selections call for.
Questions from the gate
No — honestly. DataFort is a service in formation. The practice, methods, and doctrine on this page are the real design; the client roster is not. We are holding founding-cohort briefings for a small number of organizations ahead of a 2027 launch, and founding clients will shape the standing services.
Brian Guye — AI Architect and Engineer. DataFort is founded and led by Brian Guye, a principal AI architect who designs and builds enterprise systems as an independent consultant inside client teams. His current work is production agentic AI: multi-agent orchestration, custom Model Context Protocol (MCP) tooling, retrieval-augmented generation over enterprise data, and evaluation gates for regulated clients, spanning AWS and Microsoft Azure.
That background — eight years as an architect in EMC’s Documentum consulting division, then cloud and solutions architecture across finance, defense, and healthcare, including Fortune 100 and US military clients — is the ground DataFort stands on. The same discipline that hardened enterprise content and cloud platforms now gets pointed at a newer perimeter: your models, your agents, and the data behind them.
No. Counterbattery works entirely inside your own perimeter: decoy models, honey endpoints, and tarpits that waste an attacker’s compute and feed their extraction pipelines corrupted results. Nothing we do touches systems you don’t own. It is the difference between a castle’s murder holes and riding out to burn a village — we build murder holes.
A red team visits, tests, and leaves a report. DataFort starts there — the Bastion Audit is that visit — but Rampart stays on the wall afterward, and Counterbattery answers live assaults while they are happening. Most AI attacks are campaigns, not events; the defense has to be a garrison, not a guest.
Whatever you run. Frontier APIs, open-weight models, fine-tunes, retrieval systems, and agent frameworks — the attack classes in the siege catalogue apply across all of them, and the audit is built per-stack. We sell no model and resell no platform.
An inventory of your AI systems (we will help you build it if it doesn’t exist), a test window, and one engineer who knows where the bodies are buried. Everything runs under NDA, in writing, with rules of engagement agreed before the first probe is sent.
The Bastion Audit is fixed-scope and quoted after a scoping call — the price depends on the size of the perimeter, nothing else. Rampart is a monthly retainer; Counterbattery is activated per campaign. Founding-cohort terms are set during briefings, not published, because the cohort is small by design.
Founding-cohort briefings are open now. Build your stack in the defense builder, bring the briefing request with you, and we will walk your perimeter together.
Start with the defense builderWeb résumé
Principal AI Architect | Cloud / Solutions Architect
AI Architect and Engineer. This is the short-history web edition — career highlights and current practice, without contact details.
Principal AI Architect designing and building enterprise systems as an independent consultant inside client teams. Current work is production agentic AI across a multi-vertical client portfolio, anchored by a healthcare cloud platform for a major US medical laboratory provider: multi-agent orchestration, custom Model Context Protocol (MCP) tooling, Retrieval-Augmented Generation (RAG) over enterprise data, and evaluation gates for regulated clients, spanning Amazon Web Services (AWS) and Microsoft Azure. Background spans finance, defense, and healthcare delivery, including Fortune 100 and US military clients. Publishes applied AI architecture research.
AI architecture and hands-on build across enterprise client engagements, anchored by a healthcare cloud platform.
Independent contract practice, project-based; runs alongside the principal roles above. Clients described by vertical where names sit under NDA.
A.A.S., Computer Science — El Centro College
Short-history web edition. Full history available on request through a DataFort founding-cohort briefing.
Back to DataFort