AI Systems Engineering
Agents, RAG pipelines, LLM integrations, and document automation, taken from idea to production. 20+ systems shipped end-to-end: self-hosted, Dockerised, and built to run every day, not to demo once.
I build AI systems. Then I secure them.
I’m Nick Falshaw. I’ve shipped 20+ production AI systems, agents, RAG pipelines, and document automation, and I’ve spent 17+ years securing enterprise networks in banking, automotive, manufacturing, and regulated infrastructure. Building AI and securing AI are usually two hires. Here they’re one.
LIVE
17 years, distilled
$ whoami Nick Falshaw, AI Engineer & AI Security Consultant $ cat focus.txt AI systems engineering AI security engineering Firewall automation at scale Cloud, zero trust & ISO 27001 $ cat creds.txt AI-102 · AZ-500 · ISO 27001 LI · CEH · TOGAF 9 · CCIE Sec (written) $ _
Build the system, secure the stack, prove it in audit.
Need an AI system built, or an existing one secured? I do both: agents and pipelines that ship, controls that survive production pressure, and the evidence to prove it. Firewall automation and ISO 27001 stay in the toolkit.
AI Security Engineering
Agents take action, RAG pipelines expose context, and MCP servers expand the attack surface. I threat-model agentic systems, lock down ingestion, and harden self-hosted LLM stacks before they reach production.
Firewall Automation
Vendor-agnostic change automation grounded in 280+ real migrations across Palo Alto, Check Point, Cisco, Fortinet, and F5. Rule sets that survive audit, segmentation that holds, and automation that removes the human bottleneck.
Cloud & Zero Trust
Identity-first security for Azure and hybrid estates: conditional access, segmentation, and least privilege, prioritised by the risk each control actually removes rather than the logo on the box.
ISO 27001 & Compliance
ISO 27001 from gap analysis through certification, plus NIS2 and DORA readiness for regulated sectors. Passing the audit is the easy part, the goal is a programme that still holds the day something breaks.
Proven, not slideware
Every claim here is backed by delivery: production systems, real migrations, and audit evidence from regulated environments. A track record you can probe before we speak.
20+production AI systems shipped
17+years in enterprise cyber
280+firewall migrations delivered
Seventeen years close to production.
Production firewalls, regulated audits, and AI systems shipped end-to-end, the record, not the pitch.
-
2025, now
Independent AI Engineer & AI Security Consultant
Building and securing AI systems: agents, RAG pipelines, and document automation on the build side, agentic threat modelling and self-hosted LLM hardening on the defence side. Firewall automation and ISO 27001 programmes for regulated estates.
-
2010, 2025
Senior / Lead Network Security Contractor
Fifteen years contracting into DAX-30 and enterprise environments, banking, automotive, manufacturing, payments and the public sector. Delivered 280+ firewall migrations and the security architecture behind them, multi-vendor and audit-ready.
-
earlier
Network & Security Engineering
Enterprise routing, switching and perimeter security, the grounding that seventeen years of firewall, compliance and now AI-security work is built on.
Certifications
AI-102, Azure AI Engineer
AZ-500, Azure Security Engineer
AI-900, Azure AI Fundamentals
ISO 27001 Lead Implementer
CEH, Certified Ethical Hacker
TOGAF 9
CCSP
CCIE Security (written)
CCNP
CCDP
CCSA · CCSE (Check Point)
JNCIA-FWV · JNCIS-FWV (Juniper)
Palo Alto EDU-201/205/311/121
F5 BIG-IP LTM
ITIL v3 Foundation
Three ways to bring me in.
Most engagements start as one of these and grow from there. Fixed scope where it helps, retained where the work is ongoing.
AI & security review
Threat-model an AI system or audit a network estate: agents, RAG pipelines, MCP servers, firewalls, and identity. You get a prioritised findings report with fixes ranked by the risk each one removes, not a checklist.
Build the system
Design and ship the AI workload end to end: agents, retrieval, and document automation, self-hosted and hardened before it reaches production. Built to run every day, with the security baked in rather than bolted on.
Firewall & segmentation
Vendor-agnostic firewall automation and migration grounded in 280+ real projects across Palo Alto, Check Point, Cisco, Fortinet, and F5. Rule sets that survive audit and segmentation that actually holds.
Field notes.
Practical writing on firewalls, compliance, and shipping AI systems that survive real users.
JadePuffer: Ransomware With No Human at the Keyboard
The first ransomware run entirely by an AI agent, foothold to encryption, no operator. It went from failed login to working fix in 31 seconds. What breaks when the adversary's tempo and cost both collapse.
ReadYour AI Coding Tool Was Profiling You. The Real Lesson Is About Trust.
A researcher found Claude Code quietly fingerprinting users and phoning markers home. The scandal is not Anthropic. It is that you gave a vendor agent root on your codebase and never audited the trust.
ReadThe Security Controls That Pass the Audit and Do Nothing
The unread access review, the log-only WAF, the phishing-completion metric, the rubber-stamp firewall review, the regex "guardrail". Controls that pass audit and stop nothing, and how to test effectiveness instead of existence.
ReadThe Credential Is Dead. Long Live Continuous Verification.
16 billion credentials leaked and stolen sessions walk past MFA. Possession stopped proving identity. The fix is continuous verification, at the firewall and the AI agent alike.
ReadAI Agents Are the Service Accounts That Learned to Think
AI agents are non-human identities with standing access. After 17 years cleaning up orphaned firewall rules, the sprawl, and the cure, are familiar.
ReadThe Autonomous AI Attacker Arrived. It Won in Under an Hour.
An AI agent ran a real intrusion end to end, foothold to data exfiltration, in under an hour with no human. What it breaks for defenders.
ReadTwo Authentication Bypasses, One Bug Class: The VPN Edge and the AI Stack
A Check Point VPN zero-day and the BadHost flaw in Starlette are the same CWE-287 mistake, one stack apart: trusting attacker-controlled input to authenticate.
ReadFrom CVSS to ASR: Putting a Number on AI Agent Risk
For 17 years I scored security risk in CVSS. Anthropic just gave AI agents their own number: a 31.5% prompt-injection hijack rate. The risk discipline transfers.
ReadPrompt Injection Is a Confused-Deputy Problem We Already Solved
Prompt injection is a confused-deputy attack: the agent spends its own authority for an attacker. Network security solved that with trust boundaries, not smarter parsing.
ReadThe Firewall CVE and the AI-Agent Breach Are the Same Mistake
A root-RCE firewall CVE and a poisoned AI-agent skill marketplace share one root cause: no provenance, no least privilege, no change control.
ReadOWASP LLM Top 10: 5 Critical AI Vulnerabilities for 2026
Where production LLM systems break first, based on mapping the OWASP LLM Top 10 against real codebases.
ReadZero Trust Mittelstand: A Pragmatic 90-Day Plan
Zero Trust for the Mittelstand, not the Fortune 500: identity, segmentation, and continuous verification without a full rebuild.
ReadMittelstand NIS2: Why German SMEs Are Dangerously Unprepared
The backbone of Europe’s economy built its advantage through engineering, not security. Here is what BSI deadlines require and how to close the gap.
ReadWhat 15 Years of Enterprise Security Compliance Taught Me
Hundreds of assessments across TISAX, PCI-DSS, ISO 27001, and NIS2, plus the lessons that do not appear in any framework.
ReadAI Threat Detection: 7 Strategies for CISOs in 2026
How to reduce alert noise, find novel attacker behaviour, and turn detection maturity into evidence the board can understand.
ReadHow an engagement works.
The method stays consistent whether I’m building an AI system or securing one: evidence first, architecture second, implementation third, verification always.
01
Assess
Map what’s really deployed, topology, AI workloads, identity, threat surface, regulatory scope. Facts, not assumptions.
02
Architect
Design the target state and the path to it. AI controls and network defences on one blueprint, prioritised by risk and effort.
03
Implement
Build it: firewall automation, RAG controls, identity, and segmentation hardened against the vectors that actually land.
04
Verify
Prove it under real load. Useful alerting, owned runbooks, and audit evidence produced as part of delivery.
Ship AI you can
stand behind.
Building an AI system? Running one that’s never been threat-modelled? Still on firewalls and ISO 27001? Tell me what you’re building or defending and what deadline matters. I reply within one business day.
Message sent.
You’ll get a real reply from me within one business day.