Your vendor switched on AI. Nobody told you.

You ran the vendor review. You read the data-processing agreement. You approved the tool.

None of that protects you from what happened next - the vendor quietly shipped an AI feature, switched it on by default, and began sending your data to a model you never assessed.

New research says 64% of them never told you.

That figure; vendors who failed to disclose new AI subprocessors in their data-processing agreements - is the clearest sign yet that third-party risk has quietly broken.

The tool you vetted six months ago for cloud storage or support tickets can push a "summarise with AI" feature overnight, piping your customers' data to an external model, announced only through a subprocessor-update email that never reaches the people who'd object.

And the contract you're leaning on may not save you. Plenty of these features launch with "help improve our product" sharing switched on by default. Your agreement can forbid model training in writing while the toggle in the admin console does the opposite - and defaults win until someone actively turns them off. A clean contract offers no protection against a dirty configuration.

This is fourth-party risk, and it's the part your questionnaire never reached. Your vendor may run its AI on one provider's model and host it on another's infrastructure ... sub-processors you never saw, in a chain you never mapped. When the Salesloft-Drift breach hit last year(sources in comments), the blast radius of a single compromised SaaS-AI integration reached far wider than earlier supply-chain incidents, because that one connection quietly touched hundreds of downstream environments.

Here's the shift it forces. A vendor review is a photograph; you need a film. When a supplier can switch on a new model next Tuesday, the annual re-assessment is a comforting fiction. And the fix isn't more contract language - it's evidence. Stop accepting the clause and ask for the screenshot of the console showing the training toggle is off. Tier vendors by what their AI actually touches.._not_ by what they told you last year.

For regulated firms there's a sting in the tail: under the EU AI Act and ISO 42001, if your vendor's embedded AI is high-risk, the obligation can follow you as the deployer - and moreover, insurers have started asking not whether you use AI, but how you govern it across everyone you buy from!

None of this has a settled framework yet. The third-party playbook we all trained on assumed a vendor stayed the thing you assessed - and that assumption is gone. Antonio Machado put it better than any standard does: "Wanderer, there is no path; the path is made by walking." This corner of governance is being mapped by the people walking it right now.

So, for your next board pack: you know which vendors hold your data. Do you know which ones started feeding it to an AI last quarter ... and would you find out before they did?