Last week, three different Slack threads raised similar concerns. A developer questioned if ChatGPT was a "sanctioned tool." An engineer wondered if copying code into Claude breached data handling policies. A manager needed to know the number of AI accounts in the organization before a compliance audit.
If your team is asking similar questions, you're not alone. AI tools are spreading through organizations like Slack and Dropbox did years ago. Someone finds a tool that solves a problem, shares it, and suddenly there are 200 accounts you didn't authorize.
Here's what security teams are asking and what you need to know.
What Counts as "Shadow AI"?
Shadow AI refers to any AI tool processing company data without your approval. This includes:
- Developers using ChatGPT or Claude for debugging
- Sales teams using AI writing assistants accessing your CRM
- Marketing uploading customer data to AI image generators
- Use of browser extensions that send page content to large language models (LLMs)
The issue isn't "free vs. paid" or "consumer vs. enterprise." It's whether the tool went through your vendor review process. If someone signed up with their work email and started using company data without IT approval, it's shadow AI.
This is important for ISO/IEC 27001:2022 compliance. Control A.5.23 requires maintaining information security during supplier relationships. If you don't know which AI vendors your employees are using, you can't assess those relationships.
How Do We Find These Tools?
Start with your identity provider. Most AI tools use OAuth for sign-ins, creating an audit trail.
If you're using Microsoft 365 or Google Workspace, pull OAuth grants to see which third-party apps have access. Look for patterns: multiple grants to the same domain, unfamiliar tools, apps requesting broad permissions like "read all email" or "access all files."
This process can be manual. You're exporting logs, parsing JSON, and trying to identify which "Acme AI Assistant" your marketing team mentioned. Tools like Nudge Security automate this by integrating directly with your identity provider, giving you immediate visibility into every AI app connected.
Don't stop at OAuth. Check:
- Browser extension policies (many AI tools run as Chrome extensions)
- Expense reports (monthly subscriptions show up here)
- Network logs for domains like openai.com, anthropic.com, or unusual API endpoints
- Your CASB if deployed
Do We Need to Block Everything We Find?
No. Blocking creates friction and pushes people toward workarounds. You'll end up with developers using personal email accounts or copying data to their phones.
Instead, categorize what you find into three buckets:
Acceptable with conditions: Tools you can approve if teams follow specific guardrails. Maybe ChatGPT is fine for public documentation but not for customer data. Document those conditions clearly.
Needs review: Tools that might be acceptable but require vendor assessment. Check their data handling practices, data storage locations, and whether they train models on customer inputs. This is where SOC 2 Type II reports and data processing agreements matter.
Block immediately: Tools that violate non-negotiable policies. If your PCI DSS v4.0.1 scope includes cardholder data, you can't have people pasting that into public LLMs. Block at the network level and explain why.
The key is providing approved alternatives. If you block ChatGPT for code review, point them to GitHub Copilot on your enterprise plan. If you block consumer AI writing tools, provide an enterprise option that meets your data handling requirements.
How Do We Write Policies People Will Follow?
Make them specific and example-driven. "Don't use AI with sensitive data" is too vague. Instead:
"You can use ChatGPT for:
- Drafting public blog posts
- Generating test data (no real customer names or emails)
- Explaining error messages from open-source libraries
You cannot use ChatGPT for:
- Code that processes payment data
- Customer support responses (use our approved tool)
- Anything marked Internal or Confidential in our data classification scheme"
Include screenshots showing where to check data classification labels. Link to your approved alternatives. Make it easier to do the right thing than to guess.
Review your policy quarterly. AI tools change fast. The model that didn't retain training data last month might have new features this month. Your policy needs to keep up.
What If We're Already Failing Compliance Audits?
Start with containment, then remediation. If you're in the middle of a SOC 2 audit and just discovered 50 unapproved AI tools, here's your 48-hour plan:
- Document what you found (tool names, user counts, data types accessed)
- Identify which tools touched sensitive data
- Revoke OAuth access to anything high-risk
- Brief your audit team with what you found and what you're doing about it
Then build your ongoing program:
- Set up continuous monitoring (manual quarterly reviews don't catch new tools fast enough)
- Create your approved AI tools list with clear use cases
- Build a vendor review process specifically for AI tools (they need different questions than traditional SaaS)
- Train teams on the policy with real scenarios they'll encounter
For NIST CSF v2.0, this maps to the Govern function (GV.OC-03: cybersecurity and privacy risks are incorporated into organizational risk management) and the Identify function (ID.AM-02: software platforms and applications are inventoried).
Where Do We Go From Here?
Your immediate next step: run an AI tool discovery scan this week. Use your identity provider logs, check browser extensions, and talk to team leads about what they're using. You can't govern what you can't see.
Then set up continuous monitoring. Shadow AI isn't a one-time cleanup project. New tools launch monthly, and your teams will keep experimenting. You need real-time visibility, not quarterly audits.
Finally, shift your security culture from "block by default" to "approve with guardrails." The teams using AI tools aren't trying to create security incidents. They're trying to work faster. Your job is to help them do that safely.
