Services we provide
Build secure AI from day one
Who it's for
Founders and small teams building their first AI product — a chatbot, an AI agent, an MCP integration, or an AI feature inside an existing app.
What you get
- We work with you on choosing the right model — cloud LLM (OpenAI, Claude, Gemini) or a local model running on your hardware — and on the right tools.
- A threat model written in plain English, covering prompt injection, indirect injection through RAG / logs / external data, and agent hijack.
- Secure defaults for prompts, system instructions, data handling, and API keys.
- A review of your training or fine-tuning data, with a note on what the AI Act and GDPR allow you to do with it.
- A short "before launch" checklist.
- Recommendations for ongoing tests and observability tools (so you can see what your AI is doing).
What it does not cover
The actual product design, the model fine-tuning, or building the app itself. We review and steer the security work; you (or your developer) keep building.
Find and fix risks in an existing AI system
Who it's for
Teams and founders who already have an AI product live — a chatbot, an AI agent, a RAG system, an MCP server, an internal copilot, or a combination — and want a calm, evidence-based review of what is risky.
What you get
- A full review against OWASP Top 10 for LLMs, OWASP Top 10 for Agentic Applications, OWASP MCP Top 10, and MITRE ATLAS.
- A focused look at indirect prompt injection through RAG documents, web pages, log files, and MCP tool descriptions — one of the most common real-world attack paths.
- A review of your agent permissions and observability — does your agent have more access than it needs? Can you actually see what it is doing?
- A review of your API key handling, secret storage, and service-to-service authentication.
- A check on data handling: what your AI stores, where it sends it, and how that fits the AI Act and GDPR.
- A written report with severity-rated findings and step-by-step fixes.
- A follow-up call to walk through it. Retest after fixes if you want one.
What it does not cover
Legal advice, regulatory filing, or rewriting your application code.
Test your AI like an attacker would
Who it's for
Teams and founders who want to know what a real attacker — or a curious user — could do to their AI before someone else finds out.
What you get
A written scope and written authorisation before any test. We never run a test without your permission, in writing, on a defined target, within a defined time window.
- A structured campaign of attacks adapted to your system: direct and indirect prompt injection, jailbreaks, data-leakage tests, tool and API key abuse, agent hijack, MCP tool-poisoning attempts, and supply-chain probes.
- An assessment of your agent's blast radius — what an attacker who takes over one tool call can reach.
- A report mapped to OWASP LLM Top 10, MITRE ATLAS, and (where relevant) the AI Act and GDPR risk categories.
- A retest after fixes.
What it does not cover
Anything outside the written scope; anything without your written authorisation. We do not test production systems belonging to a third party, and we do not sell exploit code or payloads to anyone.
Secure your MCP server
Who it's for
Teams connecting AI assistants (Claude Code, Cursor, Windsurf, custom agents) to internal tools, files, or APIs through the Model Context Protocol — and who want the integration audited before it goes live, or audited after a scare.
What you get
- A review of your MCP server configuration: tool definitions, scopes, allowed inputs, and the principle of least privilege.
- Checks against the OWASP MCP Top 10, with special attention to tool poisoning, prompt injection through tool output, and unauthenticated command-injection paths.
- Supply-chain vetting of any third-party MCP servers you plan to use.
- Guidance on sandboxing (running MCP servers in isolated environments), signing (proving a server is what it says it is), and pinning (locking the version you trust).
- A short report with severity-rated findings.
What it does not cover
Building the MCP servers for you, or hosting them. The deliverable is review and guidance.
Infrastructure and environment hardening
Who it's for
Teams and founders running their product in the cloud, on a VPS, or on their own hardware — and who want the environment around the app audited, not just the app itself.
What you get
- A review of your development, UAT and production environments: operating system configuration, packages and updates, user accounts and SSH keys, firewalls and routing rules, log retention and rotation.
- A check on containers and virtualisation: image hardening, network segmentation, runtime permissions, secret mounting.
- A review of supporting services: databases, caches, message queues, object storage, TLS certificates.
- Backups: are they running, where do they go, can you actually restore from them.
- A short report with severity-rated findings and step-by-step fixes.
What it does not cover
Building the environment for you, or 24/7 monitoring. The deliverable is a review and the fixes.
CI/CD pipeline hardening
Who it's for
Teams and founders who push code from a Git repository to a running product through a build and deploy pipeline — and who want that path audited before a leak or a breach.
What you get
- A review of your source-control setup: branch protection, code review rules, signed commits, secret scanning in history.
- Build hardening: pinned dependencies, checksum or signature verification for downloaded packages, isolated build runners.
- A check on the pipeline itself: who can change it, what runs in it, where the artifacts and logs go.
- Library and dependency vetting: checks against known-vulnerable and known-malicious packages, including AI-coding-tool hallucinations.
- Deployment safety: staged rollouts, rollback paths, drift detection between environments.
- A short report with severity-rated findings.
What it does not cover
Writing your pipeline for you, or hosting CI/CD runners. The deliverable is review and the fixes.
Training and consulting for AI adoption
Who it's for
Teams and institutions that want to use AI safely across the organisation — without becoming security experts themselves. This service is also for teams who want to start using AI agents but do not yet know how they work or how to keep them in check.
What you get
- Workshops for non-technical staff on AI agents, vibe coding, safe AI use, and the AI Act / GDPR basics.
- A one-page AI acceptable-use policy template you can adopt as-is or adapt.
- Training on AI observability — how to read the logs, traces, and costs coming out of Langfuse, Laminar, Helicone, or your own stack.
- Support during AI vendor evaluation: which questions to ask a vendor before you sign, and which answers are red flags.
- Guidance on local models versus cloud APIs: when a local model makes sense (sensitive data, predictable behaviour, lower cost at scale), and how to deploy one securely.
- Quarterly advisory calls. Support during onboarding of new AI tools.
What it does not cover
The actual implementation of AI features in your product, or any technical build work.