Service / 09

AI Security

AI systems add attack surface that classical security programs do not cover: prompt injection, data exfiltration through model outputs, poisoned inputs. We threat-model, harden, and test AI systems before they ship — and keep testing after.

When you need this

The situations we are called into.

You are shipping an LLM feature that reads untrusted content — documents, email, the open web.

Your security team's threat models predate language models, and nobody has updated them.

An AI system needs security sign-off before go-live, and generalist firms lack the expertise to give it.

Approach

How the engagement runs.

Threat modelling

Attack surface mapped for each AI system: inputs, tools, outputs, and the data each can reach.

Architecture review

Trust boundaries, privilege separation, and output handling reviewed against the threat model.

Adversarial testing

Injection, exfiltration, and abuse scenarios executed against the real system, with reproducible findings.

Hardening

Fixes implemented and verified — input mediation, output constraints, privilege reduction — not just listed.

Monitoring design

Detection rules and response procedures for the attacks that will arrive after launch.

Deliverables

What you hold at the end.

AI threat model

Adversarial test report

Implemented hardening measures

Detection rules and response procedures

Secure-deployment checklist

Related research

Technical report · 2026

Evaluating large language models for regulated enterprise workflows

A practical evaluation framework for LLM systems in banking, insurance, and government contexts — metrics, test harnesses, and a taxonomy of failure modes.

In preparation

White paper · 2026

Multi-agent systems in production: reliability patterns

Failure modes observed when orchestrating multiple LLM agents on business-critical tasks, and the containment patterns that keep them recoverable.