Book a consultation

AI Security

AI Security and AI Governance Consulting

We help engineering and risk teams ship AI features safely: threat modelling for LLM and agentic systems, hardening of AI applications and pipelines, data and model protection, and a governance framework that keeps AI adoption accountable as it scales.

The problem we solve

AI moved from experiment to production faster than most security programmes could adapt. Teams are shipping retrieval pipelines, agents with tool access and third-party model integrations against controls designed for conventional web applications.

The failure modes are genuinely different. Prompt injection turns untrusted content into instructions. Agents with tool or API access turn a text vulnerability into an action. Retrieval systems quietly widen who can reach which data. Model and prompt assets leak through logs, caches and vendor integrations.

Meanwhile boards, enterprise customers and regulators are starting to ask concrete questions about AI risk — what data goes to which provider, who approved this use case, how the system is monitored, and what happens when it behaves badly. Most organisations have no documented answer.

AI security is not a new discipline bolted onto the side of your programme — it is application security, data governance, identity and supply-chain security applied to a system that takes instructions from untrusted text and can act on them. We approach it that way: reuse what already works in your security programme, and add controls only where the AI architecture genuinely creates new exposure.

Our approach

Adapted to your environment and constraints — but the shape of the work is consistent.

  1. AI inventory and use-case triage

    We map where AI is actually being used — first-party models, hosted APIs, embedded vendor features, developer tooling — and classify each use case by data sensitivity, autonomy and blast radius. You cannot govern what you have not enumerated.

  2. AI threat modelling

    We threat model the specific architecture: prompt and context flows, RAG sources, tool and function calling, agent autonomy boundaries, plugin and MCP integrations, training and fine-tuning pipelines. We work from recognised references such as the OWASP Top 10 for LLM Applications, MITRE ATLAS and the NIST AI Risk Management Framework rather than a generic checklist.

  3. Control design and hardening

    We design proportionate controls — input and output handling, tool allow-listing and least-privilege for agents, tenant and document-level authorisation in retrieval, secrets and key handling, output filtering, rate limiting and abuse detection — and work with your engineers to implement them.

  4. Evaluation and continuous testing

    We help you build adversarial test suites and red-teaming exercises for AI behaviour, and wire them into CI so regressions are caught before release rather than by a customer.

  5. Governance that keeps up

    We put a lightweight AI governance framework in place: an acceptable-use policy, an intake and approval path for new AI use cases, a model and vendor register, data-handling rules per provider, human-oversight expectations, logging and monitoring requirements, and an incident pathway for AI-specific failures.

Expected outcomes

What changes as a result of the engagement.

  • A complete, maintained inventory of AI systems, models and data flows
  • Documented AI threat models with prioritised, owned remediation
  • Reduced likelihood and blast radius of prompt injection and agent abuse
  • Clear boundaries on what data may reach which model or vendor
  • Repeatable adversarial testing running in your pipeline
  • An AI governance framework that supports customer and audit questions
  • Faster approval of new AI use cases, because the path is defined

Typical deliverables

Confirmed in the proposal before work starts, and adjusted to scope.

  • AI system and use-case inventory
  • AI threat model report with prioritised findings and remediation plan
  • Secure AI architecture and reference patterns for your stack
  • AI security control standard for engineering teams
  • Adversarial test suite and red-team findings report
  • AI acceptable-use policy and use-case intake/approval workflow
  • Model, vendor and data-flow register
  • AI governance operating model with roles, RACI and review cadence
  • Executive briefing on AI risk posture

What this covers

The specific capabilities available under this service. Engagements usually draw on a subset — we scope to the problem, not the catalogue.

AI application security

  • LLM application security review
  • Agent and tool-use authorisation design
  • RAG and vector store access control
  • Prompt injection and jailbreak defence
  • Output handling and downstream injection
  • AI abuse, misuse and rate-limit controls
  • Secrets and key management for AI services

Model & data protection

  • Training and fine-tuning data governance
  • Data classification for AI pipelines
  • PII minimisation and redaction
  • Model artefact and weight protection
  • Provider data-retention review
  • Tenant isolation in multi-tenant AI

AI governance & risk

  • NIST AI RMF alignment
  • ISO/IEC 42001 readiness mapping
  • EU AI Act exposure assessment
  • AI acceptable-use policy
  • AI vendor and model risk assessment
  • Human-oversight and escalation design
  • AI incident response playbooks

Assurance

  • AI red teaming
  • Adversarial evaluation suites
  • Continuous AI security testing in CI
  • AI security awareness for engineers

Who this is for

  • Engineering leaders shipping generative AI features into production
  • CTOs and heads of platform at AI-native product companies
  • CISOs and risk leaders asked to approve AI adoption
  • Compliance teams facing customer or regulatory questions about AI
  • Product teams integrating third-party LLM APIs or agent frameworks

Recognise your situation? A 30-minute discovery call is the fastest way to find out whether this is the right engagement.

Book a security consultation

Common questions

Do you work with hosted model APIs as well as self-hosted models?

Yes. Most engagements involve a mix — a hosted frontier model for some features, an open-weight model running in your own cloud for others, plus embedded AI in SaaS tools. The controls differ, so we assess each path separately and give you one consolidated risk picture.

Is this a penetration test or a governance exercise?

It can be either, and most clients need some of both. A technical AI threat model and red team tells you what is broken now; the governance framework stops the same class of issue arriving with the next ten features. We scope them as separate phases so you can start where the pressure is.

How do you avoid slowing down AI delivery?

By making the safe path the easy path. Reference patterns, pipeline checks and a fast intake process for low-risk use cases mean most teams ship without a bespoke review, and scrutiny concentrates on the genuinely high-risk systems.

These engagements are often scoped together — the underlying risks overlap.

DevSecOps & AppSec

DevSecOps

Build security into the delivery pipeline instead of bolting it on at the end — secure SDLC, CI/CD hardening, IaC scanning, supply-chain controls and guardrails engineers will actually keep.

DevSecOps & AppSec

Application Security

Threat modelling, secure design review, application security testing and a vulnerability management process that closes findings instead of collecting them.

Advisory & Compliance

Cybersecurity Advisory

Security strategy, architecture review, maturity assessment, vulnerability management and incident readiness — built into a programme with owners, metrics and a roadmap leadership can fund.

Discuss your security challenges

Tell us what you are trying to secure and where it hurts. We will tell you what we would do first, whether or not you engage us.

Book a security consultation Explore our services