SOC 2 with AI

Is It Safe to Put Security Documents into ChatGPT or Claude?

The short answer

It can be, if you check four things first and redact before you paste. Using ChatGPT or Claude for compliance work means sending documents outside your environment, so find out exactly what leaves it, who can see it, whether it is retained or used for training, and which contract covers your plan. Remove secrets, personal data, customer data and exploitable detail from anything you upload. Then treat the AI vendor as one of your vendors.

Key takeaways

  • The answer depends on your plan, your settings and your contract. Read the vendor’s own documentation for the plan you use.
  • Redact a copy before the assistant sees it. Keep the original unchanged.
  • Never paste secrets, keys or tokens. Rotate any that were pasted by mistake.
  • An AI vendor that receives company information belongs in your vendor inventory, with a review on file.

What should we check before pasting security documents into ChatGPT or Claude?

Four questions decide whether a document may go into an assistant. Answer them once for each plan your company uses, and write the answers down.

Data-handling questions to answer before using an AI assistant for compliance work, and where to find each answer.
QuestionWhat to find outWhere the answer is
What leaves your environment?The text you paste, the files you upload, and everything a connector or agent can read on your behalfYour own configuration: which folders, repositories and accounts the assistant can reach
Who can see it?Which colleagues share the workspace or the conversation, and whether the vendor’s staff or subprocessors can access contentYour workspace settings and the vendor’s terms
Is it retained or used for training?How long conversations and files are kept, whether you can delete them, and whether content is used to train modelsThe vendor’s privacy documentation and your plan’s settings
Which contract covers it?Whether your company accepted commercial terms or an individual accepted consumer terms, and whether a data processing agreement appliesThe agreement itself, and whoever signed up

Check your customer contracts too. Some restrict where customer data may be processed or which subprocessors you may use.

Is a business plan different from a consumer plan?

Generally, yes. Business and API plans are typically covered by commercial terms and consumer plans by consumer terms, and the two can differ on training and retention. We do not summarize any vendor’s terms here: they change, and the vendor’s own page is the source. Read the page for the plan you use.

Find out which account each person actually uses. A company agreement covers the company’s accounts. It does not cover a personal account that an employee uses for work.

What should we redact before uploading a security document?

Redact a working copy and leave the original where it is. The copy is for analysis. It is not evidence: what you send to the auditor is the original record from the system, as what auditors accept from AI and audit evidence management best practices explain.

  • Secrets, keys and tokens: passwords, API keys, private keys, connection strings, session cookies.
  • Personal data: names, email addresses, HR and background-check records. Replace them with roles or placeholders.
  • Customer data: customer names, contract terms, records from production databases or support tickets.
  • Unpatched vulnerability details: open findings from scans or penetration tests, with affected hosts and reproduction steps.
  • Exact network layouts: IP addresses, hostnames, account IDs, firewall rules, diagrams with real identifiers.
  • Documents you hold under a confidentiality agreement, such as another company’s SOC 2 report.

Policies and procedures usually need the least redaction; scan reports, access lists and architecture documents need the most. The same care applies to security questionnaires answered with AI.

Prompt: plan the redaction of a document before analysis

I want to analyze a security document with you, and I need to redact it first. Do not ask me to paste the document yet.

Ask me what kind of document it is (policy, architecture description, scan report, access list, questionnaire), which systems it describes and what I want to learn from it, so that we keep what the analysis needs and remove the rest.

Then help me plan the redaction:
1. List the categories to look for in this kind of document: secrets, keys and tokens; personal data; customer data; unpatched vulnerability details; exact network layouts such as IP addresses, hostnames and account IDs.
2. For each category, give me search patterns I can run on my own machine and a consistent placeholder for each match, for example [HOST-1] or [EMPLOYEE-A], so that the document still reads correctly.
3. Tell me which details the analysis needs and which it does not. Flag every assumption you make about the document.

I will apply the redaction myself to a copy and keep the original unchanged. When I paste the redacted copy, check it once more and tell me if anything on the list still appears, without repeating the sensitive value in your answer.

List anything you assumed that I need to verify.

Does using an AI assistant make its vendor one of our vendors?

Yes. Once company information goes into an assistant, its provider is a vendor that handles your data. Add it to your vendor inventory, tier it by the data it receives, review its terms and security documentation, and record who approved its use. The steps are in vendor management for SOC 2 compliance. If the assistant is part of the service you deliver to customers, check whether its provider is a subservice organization.

Write a short rule on acceptable use: which plans are approved, what may be shared and what may not. Then follow it. AI-written policies and descriptions are acceptable only if they describe what your company actually does. The audit tests how controls operate, not how documents read.

An agent with access to your cloud account can read far more than the file you meant to show it. Give it read-only access limited to the task: see collecting SOC 2 evidence with an AI agent.

These decisions are management’s. We do not choose your AI vendor, approve your settings or write your rule. In a Zero-Platform SOC 2 audit we examine the evidence that your controls operated. That is AI-guided preparation, CPA-performed audit; the full workflow is in SOC 2 with AI.

Who does what

Who does what when you decide what may go into an AI assistant.
TaskYour team (with or without AI)Your AI assistantThe auditor
Plan and settingsDecides which assistant and plan are approved, and who may use themExplains a clause or setting you paste in; you read the sourceNo role
RedactionRedacts a copy on its own machines and keeps the originalSuggests categories, patterns and placeholders; checks the redacted copyNot involved. Examines original records, not working copies
Vendor review of the AI providerAdds the provider to the inventory, reviews its terms, records the approvalDrafts review questions from your vendor policyInspects the vendor inventory and the reviews on file
Acceptable-use ruleWrites it, approves it and follows itDrafts wording from what you tell itReads the rule and tests whether practice matches it

Frequently asked questions

Can we upload our SOC 2 policies to ChatGPT or Claude?

Policies rarely contain secrets, so they are among the lower-risk documents to work on. Still remove names, internal hostnames and anything a customer contract treats as confidential, and use a plan your company has approved.

What if someone already pasted a key or a password?

Rotate it now. Then delete the conversation if the plan allows it, and record what happened as your incident process requires.

Is a business plan always safer than a consumer plan?

Not automatically. The terms are typically different, and they can differ on training and retention. The answer for your plan is in the vendor’s own documentation and in the agreement your company accepted. Read both.

Does the auditor care which AI assistant we use?

We do not choose or approve your tools. We look at whether the provider is in your vendor inventory, whether it was reviewed as your policy says, and whether your people follow your own rule on what may be shared.

Sources & further reading

AICPA, 2017 Trust Services Criteria (with revised points of focus, 2022); SSAE No. 18, AT-C sections 105 and 205. Vendor documentation is linked above.

Related: Vendor management for SOC 2, How to scope systems for SOC 2, SOC 2 with Claude or ChatGPT.

Product names are trademarks of their owners. Auditsuisse is not affiliated with these vendors.

Sébastien Ruosch

Work with a licensed CPA firm

Auditsuisse Assurance NA PC is a licensed US CPA firm, enrolled in the AICPA Peer Review Program and authorized to perform SOC 2 examinations and issue SOC 2 reports. You prepare; we examine your evidence. See the Zero-Platform SOC 2 audit or book a scoping call.

Back to top ↑