AI Security Assessment: use AI without losing track of your data
We find the AI tools, agents and keys in your cloud, SaaS and identity data, check them against your data and contracts, and hand you a plan that fixes what matters first.
Not ready for a call? Start with the free hardening plan: four questions, then the steps that apply to you, in order.
Updated October 2026
Sound familiar?
If two or more of these are true, AI is already moving faster than your controls. You would not be alone: in Verizon's 2026 Data Breach Investigations Report, 45% of employees were regular users of AI on their corporate devices (authorized or not), up from 15%, and 67% of users were using non-corporate accounts on those devices.
- Engineers use AI every day, and some of it runs on personal accounts nobody approved.
- A customer security questionnaire asks how you govern AI, and the honest answer is that you don't yet.
- You're about to turn on Copilot or Gemini for everyone, and you don't know what's overshared.
- An AI pilot from last year still has API keys and access, and nobody owns it.
- Coding agents run on laptops that also hold production credentials.
- Health or customer data is one pasted log file away from a consumer chatbot.
Start from evidence, not a policy
- An AI committee with no single owner writes a long policy.
- Consumer AI gets banned, and usage moves to personal accounts.
- Nobody can say which AI tools hold which data.
- Approving a new tool takes months, so people stop asking.
- We inventory what's actually in use before anyone writes a rule.
- We build a paved road: approved tools as fast as the personal ones.
- Regulated data gets hard limits. Everything else moves.
- One named owner, and a short, published path to get a new tool approved.
What you'll be able to do
See the AI tools, accounts and agents in use
Know what's in use across SSO, OAuth grants, expense reports, code and laptops, including the forgotten pilots that still have access.
Know what data can go where
Four data tiers on one page, each mapped to the AI tools approved for it, so the policy answers most employee questions on its own.
Route regulated data only to covered endpoints
A target architecture matched to our reference models, so PHI and PII reach only BAA-covered services, with logging on the path that regulated data takes.
Answer AI questions with evidence
Give customers, auditors and your board real answers about AI vendors, data flows and controls, not a policy nobody follows.
How it works
- 1
Inventory and scan
We find the AI already in use: enterprise and personal accounts, OAuth grants, agents, MCP servers, API keys and abandoned proofs of concept. An automated scan of your cloud and SaaS does most of the work, and your MDM and EDR fill in the laptops.
- 2
Data and contracts
We classify the data AI touches and check vendor terms: training, retention, subprocessors, and BAAs that name the right services.
- 3
Architecture review
We place each AI system on one of the eleven models in our AI reference architectures and mark each gap as keep, simplify or defer.
- 4
Plan and handoff
You get a prioritized 30 and 90 day plan, a draft AI policy and an AI incident runbook, walked through with your team. About 60 days later we run the checks again, once, included in the fee, and show you what moved.
What you get
- Executive summary and readout
- AI inventory with owners: tools, accounts, agents, MCP servers and credentials
- AI posture scorecard: each control marked Verified, Attested, Partial, Failed, Not checked or N/A, with the evidence behind it
- AI architecture map: each system placed on a reference architecture model, gaps marked keep, simplify or defer
- 30 and 90 day plan: each Failed or Partial control assigned to 30 days, 90 days, or accepted with a reason
- Draft AI acceptable use policy and data rules
- AI vendor and contract review: training use, retention, subprocessors and BAAs
- NIST AI RMF report and framework map
Plus an AI incident runbook, and one retest about 60 days after the readout, included. Building AI into your product or hosting your own models? Optional modules add threat models, adversarial testing and a model hosting review.
How we check: evidence first
Most AI assessments are a questionnaire. Ours starts with an automated scan of your cloud and SaaS, so the scorecard shows what's actually configured, not what someone remembers. Most controls are measured by our tooling; the rest are answered by you and marked Attested. Interviews cover only what tools can't see. Every control gets one status: Verified, Attested, Partial, Failed, Not checked or N/A, and only Verified counts as met. Anything we couldn't check is marked Not checked, never as a pass. The scan reads configuration, not content: where it finds a model API key in a workload setting, it records where it is, never the key itself.
Access is where AI breaches start. In IBM's 2026 Cost of a Data Breach report, 92% of organizations that had an AI-related breach lacked proper AI access controls, and only 40% of organizations reported using access controls on AI models and data.
Your cloud
Cloud access keys, secrets in storage and images, AI services in use, Bedrock logging, guardrails and private endpoints, model API keys in workload settings, and AI budgets. AWS today; Azure and GCP through Orca.
Your SaaS
SSO and MFA, third-party and AI integrations, shadow AI apps, open file sharing, token scopes and the alerts that should fire on new AI access.
What tools can't see
Settings inside ChatGPT, Claude and Gemini admin consoles, laptop controls in your MDM and EDR, and the design reviews no tool can do.
The checks follow our 21-step hardening checklist across identity, data, model access, agents and detection.
Aligned to the frameworks your customers ask about
When a customer or auditor asks which framework your AI program follows, each deliverable has an answer.
| Deliverable | NIST AI RMF | ISO 42001 | Also covers |
|---|---|---|---|
| AI inventory and posture scorecard | Map, Measure | AI system inventory and risk assessment | Evidence for the CSA AI Controls Matrix (AI-CAIQ questionnaire) |
| AI acceptable use policy and data rules | Govern | AI policy and objectives | EU AI Act AI literacy measures, where it applies |
| AI vendor and contract review | Govern, Map | Supplier and third-party relationships | HIPAA business associate coverage |
| AI architecture map | Measure | Risk assessment | AWS scoping matrices, OWASP Top 10 for LLM Applications and for Agentic Applications |
| 30 and 90 day plan | Manage | Risk treatment | |
| AI incident runbook | Manage | Operation and incident handling |
Who it's for
Health tech with PHI
You want AI in the product and in the team, and every answer has to survive a HIPAA review.
B2B SaaS selling upmarket
Enterprise questionnaires now ask about AI, and deals stall on vague answers.
Rolling out workforce AI
You're about to give everyone Copilot, Gemini, ChatGPT or Claude and want to fix oversharing first.
Engineering on coding agents
Claude Code, Cursor and Copilot are everywhere, and you want speed without handing agents the keys.
AI Security Assessment FAQ
What is an AI security assessment?
An AI security assessment finds the AI tools, agents and credentials we can see in your cloud, SaaS and identity data and in short interviews, checks how they handle your data and what your vendor contracts allow, and maps the gaps against a target architecture. It ends with a prioritized plan to fix what matters first.
How long does an AI security assessment take?
About three weeks for the core assessment, up to six with the build and hosting modules; we confirm on the first call.
We already have a security program. Why a separate AI assessment?
AI changes what your existing controls have to cover. As the UK National Cyber Security Centre puts it, "current large language models (LLMs) simply do not enforce a security boundary." Agents act with the access people already have, AI search surfaces old oversharing, and new vendors sit in your data path. The assessment checks your current program against those changes rather than starting over.
Will this slow our engineers down?
The goal is the opposite. We look for a paved road: approved tools that are as fast as personal accounts, with hard limits only where the risk lives, such as secrets, production access and regulated data.
Do we need a fractional CISO to do this?
No. The assessment stands on its own. Some companies have their own team carry out the plan, and some have a fractional CISO run it with them.
We handle PHI. Can you help with BAAs for AI tools?
Yes. We review which AI services your BAAs actually cover, how they need to be configured, and whether your gateway or tooling takes traffic outside that coverage.
What do we need to prepare?
A list of the AI tools you know about, read-only access to your cloud and main SaaS consoles for the scan, and an hour or two with the people who own engineering, IT and legal.
What happens to our data?
Our access is read-only, and when we find a secret we record where it lives, never its value. Scan outputs, exports and interview notes stay in IOmergent's encrypted store for your company, never on personal devices or consumer AI tools. We use AI to draft summaries only inside our own AWS account (Amazon Bedrock, no training on your inputs), a person reviews every sentence, and you can opt out. We delete your evidence 90 days after the readout or the retest, or sooner if you ask, unless you keep working with us, and we remove our read-only access at the end.
Which frameworks does the assessment align to?
The assessor scores the NIST AI RMF 1.0 (with its Generative AI Profile) using the scan evidence and interviews. The deliverables also map to ISO/IEC 42001:2023, the CSA AI Controls Matrix and its AI-CAIQ questionnaire, the OWASP Top 10 for LLM Applications and for Agentic Applications, MITRE ATLAS, the EU AI Act where it applies, and your existing SOC 2 or ISO 27001. The mapping is at category level: good for questionnaire answers and an appendix, not a certification audit. If you're working toward ISO/IEC 42001, the assessment is a gap analysis you can build on. Our AI governance cheatsheet compares the frameworks.
Read before you call
Start with a 30-minute call
Tell us which AI tools your teams use and what data they touch. We'll tell you what the assessment would cover and what you could fix before it starts. No sales pitch.