Start with the data path, not the model
Every AI system takes an input, adds context, sends it to a model, and does something with the answer. Security lives in the arrows: what goes in, what the model can reach, and what happens to the output. The five patterns below differ only in those arrows.

Five patterns and where the data goes
Assistant
Embedded copilot
Retrieval over company files
Agent
API integration
Amber boxes are where control is lost: a vendor you have not contracted with, a scope wider than the task, or an action with no approval. The audit checks each amber box for every AI system in use.
The data path inside one request
Input
Who can send what? Is regulated data stopped at the boundary?
Context
What can the model reach? Is the scope wider than the task? Could the context be poisoned?
Model
Whose model, where, retained how long, used for training?
Output
Is it validated before anything runs it? Is it logged?
Use
Does a person approve high-impact actions?
What breaks in each pattern
| Pattern | Example | What can go wrong | What the audit checks |
|---|---|---|---|
| Assistant | ChatGPT, Claude, Gemini in a browser | Staff paste client data into a consumer account; retention and training terms are the defaults | Account type, SSO, DLP at the upload boundary, browser extensions |
| Embedded copilot | Microsoft 365 Copilot, Copilot in Edge, Gemini in Workspace | Copilot surfaces files the user could always open but never found | Sharing links, sensitivity labels, restricted search, Copilot agents |
| Retrieval over company files | A chatbot over SharePoint, a help desk bot over tickets | One overshared folder becomes everyone's answer; poisoned documents steer replies | Index scope, permissions trimming, source citations, content filters |
| Agent | Copilot Studio agents, Power Automate with AI, custom agents | The agent acts with more authority than the person who asked | Agent identity, tool permissions, approval gates, logging, a stop switch |
| API integration | A workflow sending records to a model and using the output | Keys in code, no validation of output, data sent to the wrong region | Key storage, output validation, data classification, region and retention |

Why this matters in 2026
A page from the sample report
Every finding in the real report has this shape: domain, finding, severity, evidence, fix. See the full sample report.
Guided by Ali Hassani, CISO
Certified CISO (C|CISO) and CISSP with 25+ years in IT, cybersecurity, compliance and infrastructure, and dozens of Microsoft and Cisco certifications. Every AI Security Audit is read-only, evidence-based, and written so an owner and an administrator can both act on it.
Questions business leaders ask
Is a private model safer than a public one?
Not by itself. A private model moves the hosting risk to you: patching, access, logging and evaluation become your job. It helps most when a contract or regulation requires data to stay in one place.
Does "not used for training" mean the vendor keeps nothing?
No. Training use, service retention, abuse monitoring, backups and legal holds are separate terms. The audit reads the contract for each AI vendor and records all five.
What is the difference between a copilot and an agent?
A copilot answers inside an app using the user's permissions. An agent can plan several steps and take actions on its own identity or the user's. Agents need approval gates and a stop switch; copilots need sharing hygiene.
Implementation partner
OC Security Audit stays independent and read-only. If you want hands-on administration, IT Perfection can implement the fixes: DLP rules, Conditional Access, app blocking, agent identities, and ongoing Microsoft 365 support.
Not sure which patterns are running in your business?
The AI Security Audit inventories every AI system in use and classifies it by pattern, owner and data type.
Read-only audit. No changes to your systems. Irvine, Orange County and Southern California on-site; remote nationwide.