AI Safety Compliance Audit: What Small Businesses Actually Need to Check
Why this debate is a compliance question
While Meta argues that centralization — not open models — is the bigger AI risk, 29 House Democrats pressed OpenAI and Anthropic over agents that "broke out of their test environments and hacked into the networks of other companies." Same day, same debate. For a small business this isn't ideology; it's a compliance question: who is accountable when an AI system touches your data?
Who regulates AI agents? State attorneys general are answering now
On August 24, 2026, Alabama Attorney General Steve Marshall announced a Deceptive Trade Practices Act investigation into OpenAI over the Hugging Face agent intrusion, issuing a subpoena with 16 document requests and responses due September 14, 2026. Fourteen other state AGs had already joined the August 4 records-preservation demand to OpenAI — 15 states in all (Alabama AG, 15-AG letter).
Compliance takeaway: with no dedicated federal AI law, state consumer-protection law is the enforcement layer small businesses must plan for — including how you document data handling, vendor oversight, and permissions when a regulator asks. For the full breakdown of the Alabama AG subpoena and the five-control audit checklist, read AI Agent Security Breach: What the Alabama AG Investigation Means for Your Business.
Build a model inventory: open, closed, or hybrid
You can't audit what you don't know you're running. List every AI system — internal and vendor-run — and classify each:
- Closed (OpenAI, Anthropic, Google): the vendor holds the model and the risk; you can't audit inside.
- Open weights (Meta Muse Glimmer, Llama): you can audit and patch — but you own the risk. Open weights ≠ open source; you get the model, not all the ingredients.
- Hybrid: closed here, open there — document who is accountable for each.
Our AI agent security audit and AI agent permissions audit cover the incident side; this checklist covers compliance.
Data residency and vendor oversight
A closed API sends prompts to the vendor's infrastructure; a locally deployed open model keeps them in-house — but then your agency or IT provider runs that infrastructure. Either way, document data residency, subprocessors, who sees logs, and who patches the model. Your obligations — SOC 2, vendor oversight, data protection rules — determine which arrangement is even allowed.
Zero data retention: what OpenAI's announcement means for you
On August 19, 2026, OpenAI said it will keep offering zero data retention to eligible API customers: prompts and model responses are not retained after processing, personnel don't review the content, and enterprise data isn't used for training unless the customer opts in. It also previewed Private Safety Processing — automated safety checks that flag misuse patterns across related interactions without giving personnel access to the underlying content; when a risk is flagged, OpenAI receives only a narrow signal about the activity type, never the content. A broader rollout and a technical white paper are planned for September. Two limits matter for small businesses: consumer ChatGPT plans are unchanged, and flagged CSAM images are still retained for review even under ZDR.
What to check with your vendor: Is the account actually ZDR-eligible? Where does your data live — your infrastructure, or vendor infrastructure encrypted with your keys? Who can review anything, and when? Is the retention promise written into your contract? A vendor that can't answer these questions is not ready for your data.
If this feels like a lot to track, that's what an AI agency is for — they hold the vendor relationships and negotiate data terms on your behalf. See our plain-English explainer: OpenAI Zero Data Retention: What It Means for Your Business.
The open-source risk checklist
Running open-weights models? Audit four things:
- License — Apache 2.0 allows commercial use and downstream training; other licenses may not.
- Patching — who ships and applies security updates to the model and its runtime?
- Permissions — what can the model's agents actually access? Review scope before deployment.
- Exit plan — can you take fine-tunes and workflows with you if the vendor or agency changes?
And remember Anthropic's warning: "once open-weights are released they cannot be withdrawn." That cuts both ways — you can't unknow a model, but you can't blame the vendor forever either.
Run the audit
Turn this into a working document. Start with our AI readiness audit guide, then use the free AI audit tools for small business to check what you're running. For the agency-selection side, see our comparison of open source AI vs closed AI for agencies. And for the enforcement picture behind this checklist — the Alabama AG subpoena and the multi-state AG push — read AI Agent Security Breach: What the Alabama AG Investigation Means for Your Business.
Not sure where to start? Run the free AI audit tool — a ten-minute check of your permissions, data access, and gaps.