AI Security Checklist: 60 Checks Across 8 Domains
AI Security Audit · Page 13 of 15

AI Security Checklist: 60 Checks Across 8 Domains

Inventory, identity, data protection, Copilot and SaaS, agents, logging, vendors, policy. Printable.

AI Security Checklist: 60 Checks Across 8 Domains
AI Security Audit · Page 13 of 15

Self-assess in an afternoon

Most IT teams can answer forty of the sixty checks from the admin centers they already open. The other twenty are where the audit earns its fee.

AI Security Checklist: 60 Checks Across 8 Domains
AI Security Audit · Page 13 of 15

Healthcare, CPA and law firms: start with domain 3

Regulated data in AI tools is the fastest way to a reportable incident. Data protection comes first for you.

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

How to use it

Mark each check Yes, No or Unknown. Unknown is the most useful answer: it is where discovery starts. Each domain links to the page that explains it.

How to use it

The checklist

62 checks. Mark each Yes, No or Unknown.

Domain 1 · AI inventory and shadow AI

Read the page
  • A written inventory of AI apps, copilots, agents and integrations exists and was updated in the last 90 days
  • Cloud app discovery is enabled and the generative-AI category is reviewed monthly
  • Firewall or DNS logs are searched for AI domains at least quarterly
  • Browser extensions are inventoried; AI assistants need approval
  • OAuth consents to AI apps are reviewed quarterly
  • Finance exports are checked for AI subscriptions
  • Each inventory entry has an owner, a plan type and a decision (approve, tolerate, block)
  • An approved-AI catalog is published to staff

Domain 2 · Identity and access for AI

Read the page
  • Approved AI apps use SSO; consumer accounts are not used for company data
  • MFA and Conditional Access cover sign-ins to AI apps
  • No AI system runs under a shared user or admin account
  • API keys for AI services are in a vault, scoped and rotated
  • Service accounts for AI have no interactive sign-in and are monitored
  • Agents have their own identities with least-privilege permissions
  • Conditional Access policies include workload and agent identities
  • Access to AI admin settings is limited and logged

Domain 3 · Data protection and DLP

Read the page
  • Sensitivity labels exist for Confidential and Highly Confidential
  • The top sensitive sites and libraries are labeled
  • Endpoint DLP includes a group for unsanctioned AI apps
  • Uploads and pastes of labeled data to unsanctioned AI are blocked or require justification
  • DLP rules cover ePHI, card numbers, client identifiers and secrets as applicable
  • Browser policy limits paste and upload to unsanctioned AI sites on managed devices
  • Approved AI vendors do not train on your data, per contract
  • Retention and deletion terms for each AI vendor are recorded

Domain 4 · Microsoft 365 Copilot and SaaS AI settings

Read the page
  • Sites shared with "Everyone except external users" are listed and reviewed
  • "Anyone with the link" sharing is restricted or expiring
  • Restricted SharePoint Search or restricted content discovery protects sensitive sites
  • Labels that exclude content from Copilot are applied where needed
  • Copilot licenses are assigned by role, not to everyone
  • Who can create and publish Copilot agents is limited
  • AI features inside other SaaS apps (CRM, HR, finance) are inventoried and configured
  • Copilot web grounding and connectors are set deliberately

Domain 5 · Agents, connectors and MCP servers

Read the page
  • Every agent has a named owner and a written allowed-tool list
  • Agents cannot send, pay, delete or grant access without a named human approval
  • Transaction, spend and runtime limits are set
  • Agents can be paused and their identities revoked; this has been tested
  • MCP servers and plugins are inventoried and from known publishers
  • Agents never hold production credentials in prompts or memory
  • Inter-agent messages are authenticated
  • Agents built by people who have left are disabled

Domain 6 · Logging, monitoring and incident response

Read the page
  • Alerts fire for new generative-AI apps above a risk threshold
  • Alerts fire for labeled uploads to unsanctioned AI
  • Agent tool calls are logged with the requester
  • Copilot interactions are in the audit log and retained per policy
  • Admin changes to AI settings are logged and tied to change tickets
  • The incident response plan names AI systems and prompt injection
  • A tabletop exercise has covered an AI incident
  • Logs are kept for at least 12 months

Domain 7 · Vendor and contract review

Read the page
  • Each AI vendor's training-data terms are recorded
  • Retention, deletion and backup terms are recorded
  • Subprocessors and data location are known
  • Breach notification terms exist
  • BAAs are in place for any AI touching ePHI
  • Security attestations (SOC 2, ISO 27001) were obtained where available

Domain 8 · Policy, training and governance

Read the page
  • An acceptable-use policy covers what may be entered, shared, automated and published
  • Staff have acknowledged the policy in the last 12 months
  • A request path for new AI tools exists with a response-time promise
  • A risk register lists AI findings with owners and due dates
  • AI used in hiring or employment decisions is flagged and reviewed with counsel
  • A quarterly review of inventory, risks, vendors and policy is scheduled
  • An evidence file is ready for insurers, clients and auditors
  • Leadership receives a short AI risk update each quarter
AI security checklist
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

Can I download this?

Print this page, or request the Excel version with the audit. We do not gate it behind a form.

How is this different from the audit?

The checklist tells you what to look at. The audit looks, collects the evidence, rates the risk and tells you what to fix first, in a report you can hand to leadership or an insurer.

Which domain matters most for a small business?

Domain 1 and domain 3. Know what AI is in use, and keep regulated data out of the wrong tools.

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

Too many Unknowns?

That is what the AI Security Audit is for. Read-only, one fixed price, evidence for every line.

Read-only audit. No changes to your systems. Irvine, Orange County and Southern California on-site; remote nationwide.