One-time engagement
AI security review
A one to two week review of your AI system, tested for the ways it can be turned against you, from prompt injection to one customer’s data reaching another, that ends in a ranked fix list and the answers for your buyer’s security questionnaire.
Find out what your AI can be made to do before someone else does.
- 1 to 2 weeks
- From $1,500
- Ranked fix list, questionnaire answers
Is this for you?
Your AI touches customer data or money
It reads their documents, acts in their account or moves something of value. A mistake is a breach, not a bad answer.
A buyer sent a security questionnaire
Procurement wants written answers about your AI, and nobody on the team has tested the things they ask about.
Your agents use tools
They call APIs, read files or act in other systems, and you want to know what they can be talked into.
Not this? The production readiness audit scores security as one of six gates; the architecture review covers scale and cost rather than security.
What you get
- 01
The findings
Each of the five tests, pass or fail on your system, with the evidence: the input that got through and what it reached.
- 02
The ranked fix list
Every finding ranked by what it could cost you, with the fix, written for your team to act on without me.
- 03
The questionnaire answers
The answers your enterprise buyers will accept on security and data handling, written from the system as it is, with any gaps named.
Illustrative pages. Yours carry your system’s scores, numbers and fixes.
Security questionnaire
Ranked fix list
- 1CostFix
- 2CostFix
- 3CostFix
- 4CostFix
- 5CostFix
Security review findings
Five tests, run on your system
- Prompt injection
- Data leakage
- Agent permissions and tool scope
- Secrets
- The software supply chain
What I test
Five ways an AI system gets turned against its owner. I try each one on your system, on staging, with your knowledge, and record what got through.
01Prompt injection
I put instructions where your agent reads: in a customer message, a document, a web page. Then I see whether it follows them instead of yours.
I measureWhich inputs got through, and what they made the agent do
02Data leakage
I ask as one customer and try to reach another’s data through retrieval, memory and the tools the agent uses.
I measureEvery path tested, each as pass or fail
03Agent permissions and tool scope
I map what each agent and tool is allowed to do, then try to talk the agent into doing more than that.
I measureWhat each tool can touch, and what it could be made to touch
04Secrets
I look for keys and tokens where they should not be: in prompts, logs, code, traces and the places an agent can read.
I measureEvery secret found out of place, and where
05The software supply chain
I check what your dependencies, packages and build pipeline let in, and how a bad package would reach production.
I measureThe dependencies and pipeline steps that need hardening
PolinRider Scanner, the tool I built after cleaning up a real incident
1 to 2 weeks, written report and readout
The two weeks
- Days 1 to 2
Read and map
Access in place, the code and the agent and tool configuration read, every input and every tool mapped.
- Days 3 to 7
Test
The five tests run against staging, each attempt and its result recorded with the evidence.
- Days 8 to 10
Rank and write
Every finding ranked by what it could cost you, the fix beside each, the questionnaire answered, a readout with your team.
The next step
Start with a
20-minute call.
Tell me what the AI touches and who is asking about it. You leave knowing whether the review fits, how long it would take and what it would cost.
Bring
- What the AI reads and what it can do
- Who is asking: a buyer, a board, your own team
- The questionnaire, if there is one
Why me
Security has been part of the job for a long time. In 2017 I was software architect, then acting CTO, at HumanFirewall, a phishing-simulation and security-awareness platform. At Devnagri AI I led security reviews and hardening of the platform as fractional CTO, and at TWINii I led organisation-wide security and supply-chain hardening as Chief AI Architect. PolinRider Scanner, my open-source tool, came out of cleaning up a real incident.
The track record“The guy thinks 5 steps ahead. He bridges DevOps rigor with AI workflows like it's second nature. Clean infra, predictable deploys, zero drama.”
Kamaljit SinghFounder, IndraOps TechnologiesAsked on most first calls
The questions.
How long does the review take?
One to two weeks from the day access is in place. One agent with a few tools is usually one week; several agents, many tools or several customer tenants usually two.
What access do you need?
Read access to the code, the agent and tool configuration and the prompts as they run, a staging system I can test against, and your buyer’s questionnaire if you have one. I never run tests against production without your say.
Do you fix what you find?
The review ends in a ranked fix list your team can act on. Where a fix is a line or a setting I say exactly what to change. If you want me on the fixes, I can join part-time after the review.
Will you break anything?
No. The tests run against staging, with your knowledge, and nothing is changed in your system. Where I need to try an input against production I ask first.
Do you sign NDAs?
Yes, yours or mine, before I see anything.
What if you find nothing serious?
Then you have a written review that says so, with the evidence, which is what your buyer was asking for.
What does it cost?
From $1,500. The price moves with the size of the system, so it is fixed on the first call, in writing, before anything starts.
How is this different from the production readiness audit?
The audit scores six gates, security among them, over three to four weeks. The security review goes deeper on that one gate in one to two weeks. If security is the only question, this is the faster route.