AI security for what you build with AI.

Before you ship, and after.

Who we assist

Small businesses and solo founderswho have built (or are building) with AI — using AI tools to build their product, or shipping an AI feature inside it.
Internal teamswho need an AI solution or training for a specific business problem.
Educational institutionslooking for expertise in AI for training or internal use.

What we cover

AI chatbots and copilots: Customer-facing assistants, internal copilots, RAG systems.
AI agents and agentic systems: Agents that read, write, and act across your tools (including Claude Code, Cursor, custom agents).
MCP servers: The connections between AI assistants and your files, APIs, and data.
Local and cloud LLMs: Running models on your own hardware or via third-party APIs, configured safely.
API keys and secrets: The credentials that let your AI tools talk to your services.
AI Act, GDPR, and data handling: What your AI is allowed to do, what it stores, and what it must not do.
Infrastructure and environments: The systems your app runs on — OS, packages, firewalls, logs, backups, containers, and the services around them.
CI/CD pipeline: From a commit to a running product — source control, build, test, deploy, and checks for known-bad libraries and exposed secrets.

What we do

Build secure AI from day one

We work with you on model and tool choices — cloud or local — and set secure defaults for prompts, data, and API keys. A short "before launch" checklist and an AI Act / GDPR review are included.

See this service: Build secure AI from day one

Find and fix risks in an existing AI system

A calm review of your live AI system — chatbot, agent, RAG, MCP, or copilot — against OWASP and MITRE ATLAS. We look for indirect prompt injection through RAG, logs, and external data. You get a written report with severity-rated findings and step-by-step fixes.

See this service: Find and fix risks in an existing AI system

Test your AI like an attacker would

A written scope, prompt-injection and tool-abuse campaigns (including indirect injection and agent hijack), and a report mapped to OWASP and MITRE ATLAS. Requires written authorisation before any test.

See this service: Test your AI like an attacker would

Secure your MCP server

A review of your Model Context Protocol setup against the OWASP MCP Top 10. We check tool definitions, scopes, and the supply chain of any third-party MCP servers. Includes sandboxing and signing guidance.

See this service: Secure your MCP server

Infrastructure and environment hardening

A review of your development, UAT and production environments — OS, packages, firewalls, logs, containers, and the services around them. We check the same controls on every layer, from the bare host up.

See this service: Infrastructure and environment hardening

CI/CD pipeline hardening

A review of the path from a commit to a running product: source control, build, test, deploy, and the checks in between for known-bad libraries and exposed secrets. We also look at hallucinated package names from AI coding tools.

See this service: CI/CD pipeline hardening

Training and consulting for AI adoption

Workshops for non-technical staff on AI agents, vibe coding, and safe AI use. A one-page AI acceptable-use policy template, vendor evaluation support, observability training (so you can see what your AI is doing), and quarterly advisory calls.

See this service: Training and consulting for AI adoption

Book a meeting (opens in a new tab)

Security facts about AI

About 45% of AI-generated code fails basic security checks.
Almost half of what AI tools write doesn't pass simple security tests.
Source: Veracode, 2025
Around 1 in 5 packages your AI coding tool suggests don't exist.
Across 16 popular code models, 19.7% of recommended packages were fabrications. Attackers are already registering these fake names and shipping malware under them.
Source: SecurityWeek, April 2025 Full report (the USENIX paper by Spracklen et al.)
Of the AI-built apps with no login, around 40% were exposing sensitive data.
RedAccess found 5,000+ apps on Lovable, Base44, Replit and Netlify with no authentication. About 40% of them leaked medical records, financial data, or corporate strategy.
Source: Wired, May 2026
Prompt-injection attacks surged 340% year over year in 2026 — the fastest-growing category OWASP tracks.
No other type of AI attack that OWASP follows grew faster in 2026.
Source: OWASP State of Agentic AI Security and Governance 2.01, June 2026
73% of audited production AI deployments have prompt-injection weaknesses.
That is nearly three out of four AI systems that were audited while running in real use.
Source: Cisco State of AI Security 2026

How it works

Book a meeting.A short call to understand what you've built and what worries you.
We agree on a written scope.You know exactly what we'll do, what we'll touch, and what we'll deliver.
You get a plain-English report.Severity-rated findings, step-by-step fixes, and a follow-up call to walk through it.