How AI Systems Work: A Security Guide for IT Teams
AI Security Audit · Page 2 of 15

How AI Systems Work: A Security Guide for IT Teams

Assistant, copilot, retrieval over your files, agent, API integration. Five patterns, five data paths, five ways to lose control.

How AI Systems Work: A Security Guide for IT Teams
AI Security Audit · Page 2 of 15

The model is the smallest part of the system

The prompt, the connected data, the identity, the tools and the logs decide the risk. A good model does not fix a bad connector.

How AI Systems Work: A Security Guide for IT Teams
AI Security Audit · Page 2 of 15

Know which pattern you bought

A vendor saying a product "uses AI" tells you nothing. Ask which pattern it is and what it is allowed to touch.

25+ yearsIT & cybersecurity experience
Certified expertsCISSP · C|CISO · Microsoft · Cisco
Southern CaliforniaLocal and on-site, based in Irvine
NationwideRemote audits & advisory
Schedule a consultation

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.

Start with the data path, not the model

Five patterns and where the data goes

Assistant

Usertypes a question
Model (vendor)answers; vendor retains per plan
Usercopies the answer out

Embedded copilot

Userasks inside Word, Outlook, Teams
Graph searcheverything the user can open
Modelanswers with those files
Useracts on the answer

Retrieval over company files

Userasks a chatbot
Indexfiles inside its scope
Modelanswers from matched passages
Usertrusts the citation

Agent

User or triggersets a goal
Modelplans the steps
Tools and connectorsmail, CRM, scripts, payments
Actionhappens, with or without approval

API integration

Your appsends records
Model (API)returns output
Your appuses the output as data or code

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

Inputprompt, pasted text, uploaded file
Context addedsystem prompt, retrieved files, memory, tool results
Modelvendor or private
Outputtext, code, a tool call
Usea person reads it, or a system runs it

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

PatternExampleWhat can go wrongWhat the audit checks
AssistantChatGPT, Claude, Gemini in a browserStaff paste client data into a consumer account; retention and training terms are the defaultsAccount type, SSO, DLP at the upload boundary, browser extensions
Embedded copilotMicrosoft 365 Copilot, Copilot in Edge, Gemini in WorkspaceCopilot surfaces files the user could always open but never foundSharing links, sensitivity labels, restricted search, Copilot agents
Retrieval over company filesA chatbot over SharePoint, a help desk bot over ticketsOne overshared folder becomes everyone's answer; poisoned documents steer repliesIndex scope, permissions trimming, source citations, content filters
AgentCopilot Studio agents, Power Automate with AI, custom agentsThe agent acts with more authority than the person who askedAgent identity, tool permissions, approval gates, logging, a stop switch
API integrationA workflow sending records to a model and using the outputKeys in code, no validation of output, data sent to the wrong regionKey storage, output validation, data classification, region and retention
How AI systems work

A page from the sample report

Sample report · Calder Ridge Partners (fictional company) · illustrative data
Domain
Retrieval over files
Finding
Help desk chatbot indexes an HR site with 1,240 files
Severity
High
Evidence
The chatbot's search index includes the HR SharePoint site because the index was scoped to the whole tenant. Permission trimming is on, but 310 of the files are shared with "Everyone except external users".
Fix
Scope the index to the sites the bot needs, fix the sharing on the 310 files, and add the HR site to the audit's quarterly sharing review.

Every finding in the real report has this shape: domain, finding, severity, evidence, fix. See the full sample report.

Ali Hassani, CISO, OC Security Audit
CISO-led advisory

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.

IT Perfection managed IT

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.