Last updated
What is a secure AI architecture review?
A secure AI architecture review is an analytical assessment of how your AI system is designed and built, focused on where it can be attacked. Ironbark Cyber reads your architecture, data flows and key code paths, builds a threat model, and hands back a prioritised list of design changes. It's the earliest and cheapest place to remove AI risk, before a single attacker, or a single auditor, gets involved.
What does the review cover?
- Trust boundaries. Where untrusted input enters the system and how far it can travel before something trusts it.
- Data flows and context. What data reaches the model's context window, from where, and whether that respects tenant and permission boundaries.
- Tool and permission design. What the model can invoke, with whose privileges, the difference between a helpful agent and a confused deputy.
- Retrieval and knowledge sources. How RAG content is ingested, trusted and isolated, and whether it can be poisoned.
- Output handling. Where downstream code trusts model output, and what injection that opens.
- Secrets, isolation and monitoring. Key management, sandboxing of tool execution, and whether you'd detect an attack in progress.
How does it work?
We start with a scoping call and a request for design docs, data-flow diagrams and access to key code and configuration. We then run structured threat-modelling sessions with your team, referenced to the OWASP LLM Top 10, MITRE ATLAS and NIST AI RMF, and deliver a written threat model and remediation roadmap. Where the system is already running, we'll usually recommend following the review with an LLM penetration test to prove the fixes hold.
What do we receive?
- A written threat model: trust boundaries, data flows, attacker objectives and control gaps.
- A prioritised list of design and configuration changes, ranked by risk reduction.
- Framework mapping so the work feeds directly into ISO/IEC 42001 or AU AI Technical Standard alignment.
- A debrief session with your engineering and product teams.
What does it cost?
Fixed-fee, indicatively from AU$12,000, quoted exactly within one business day of a scoping call.
FAQ
Frequently asked questions
How much does a secure AI architecture review cost?
Architecture reviews are fixed-fee, quoted within one business day of a scoping call. Cost depends on the size of the system and how many integrations and data flows are in scope. Indicative pricing is published on this page.
When should we do an architecture review, before or after building?
Before, if you can. A design-stage review catches expensive mistakes, over-privileged tools, untrusted data reaching the context, weak tenant separation, while they are still cheap to fix. But a review of an existing system is just as valuable, and we do both.
What is the difference between this and an LLM penetration test?
An architecture review is analytical: we read the design, data flows and code and reason about what could go wrong. A penetration test is empirical: we attack the running system to prove what does go wrong. Reviews are cheaper and earlier; pentests are proof. Many clients do a review at design time and a pentest before launch.
Do you provide a threat model?
Yes. The core deliverable is a threat model of your AI system, trust boundaries, data flows, attacker objectives and the controls that should sit between them, plus a prioritised list of design changes.
Which frameworks do you reference?
Reviews reference the OWASP Top 10 for LLM Applications, MITRE ATLAS, the NIST AI RMF and, where relevant, ISO/IEC 42001 and the Australian Government AI Technical Standard, so the recommendations line up with the assurance work that follows.
Contact
Talk to us
Tell us what you're trying to protect, secure or build. We'll come back with a plan.