Key Takeaways
- An AI sub-processor is any company your AI vendor relies on to run models, host data or store logs on your behalf.
- The Fireworks AI breach exposed Notion AI chat transcripts even though a zero data retention commitment was in place.
- Zero data retention applies to specific endpoints and features, not to a vendor as a whole, so audit feature by feature.
- Use five steps: map the chain, trace each data flow, contract down the chain, minimise what goes in, rehearse a vendor breach.
- Under the PDPO, you remain responsible for personal data you hand to processors, however many layers deep it travels.
On 2 November, a large productivity platform learned that an AI vendor it used to run open-weight models had been breached. Within a day it had cut the connection and deleted its data from the vendor's systems. By then, chat transcripts from a small share of customer workspaces had already been exposed.
Most of those customers had never signed anything with that AI vendor. Many had never heard its name. That is the question this article answers: how do you govern the companies behind your AI vendors, when your contract only names the one in front?
What is AI sub-processor risk?
AI sub-processor risk is the exposure created when your AI vendor passes your data to other companies to run models, host infrastructure, analyse usage or store logs. Each layer adds a party you did not choose, a contract you have not read and a breach that can reach your data.
A typical enterprise AI tool now sits on a short but deep chain. The software you license calls a model provider or an inference platform. That platform runs on a cloud provider. Usage analytics, evaluation datasets and support logs may go to further services again.
Traditional vendor risk reviews were built for SaaS, where data usually stays inside one provider's environment. Generative AI changes that in two ways. First, prompts and outputs carry far richer business context than form fields: draft contracts, customer complaints, pricing logic. Second, AI vendors switch model providers often to manage cost and quality, so the chain changes faster than annual reviews can follow.
The governance gap is real. A Gartner release on 9 November 2026 reported that 71% of organisations expect to deploy agentic AI within 12 months, while only 21% have agentic AI governance capabilities in place.
What happened in the Fireworks AI and Notion breach?
An attacker reached Fireworks AI, an inference provider, through a third-party cloud provider hosting its research machines in late October. Notion, a Fireworks customer, then disclosed that AI chat transcripts from fewer than 1% of its workspaces were exposed, despite a zero data retention commitment.
According to TechCrunch's report of 12 November, this was Fireworks' second breach in as many months. Fireworks says the intrusion began at a cloud provider on 23 October and reached its development machines the next day.
In its customer notice, Notion said the exposed transcripts were used for its Insights feature and that data was retained from the Fireworks endpoints it used, even though zero data retention had been agreed. Notion had disconnected from Fireworks and deleted its data there by 3 November.
Three lessons for enterprise buyers
- The breach started two layers away from the customer: a cloud provider, then an inference vendor, then the software platform.
- Zero data retention held for some products and not for one feature that stored data by design.
- The fastest protection was operational: the ability to cut traffic and delete data within a day.
The same fortnight brought a breach at automation platform Apify, where SecurityWeek reported that one compromised credential led to further credentials, including sign-in tokens linked to Google and GitHub accounts. AI tools that connect to your systems hold exactly these kinds of keys.
How do you audit the vendors behind your AI vendors?
Use a five-step framework: map every party in the chain, trace where each feature sends and stores data, push contract terms down the chain, minimise what enters the tool, and rehearse a vendor breach. Run it for each AI tool that touches personal or confidential data, starting with the highest volume.
Step 1: Map the chain
For each AI tool, ask the vendor for its current sub-processor list and identify four roles: who runs the model, who hosts the infrastructure, who stores logs and analytics, and who supports the service. Record the country for each. If a vendor cannot produce this list within a week, treat that as a finding in itself.
Step 2: Trace each data flow, feature by feature
Zero data retention is usually granted per endpoint or per product, not for the whole vendor. Ask which features store prompts or outputs: admin dashboards, usage insights, conversation history, fine-tuning, evaluation sets and support tickets are the usual suspects. The Notion case turned on exactly this distinction.
Step 3: Contract down the chain
Require flow-down clauses so every sub-processor is bound by the same security, retention and deletion terms you agreed with the primary vendor. Add advance notice of new sub-processors with a right to object, breach notification within a fixed number of hours, and a written certificate of deletion on exit.
Step 4: Minimise what goes in
Classify data before it reaches any AI tool. Keep credentials, identity numbers and unreleased financials out of prompts, use masking where the workflow allows it, and give AI tools the narrowest system access that still lets them work. What never enters the chain cannot leak from it.
Step 5: Rehearse the vendor breach
Write a one-page runbook for the day a vendor's vendor is breached: who receives the notice, who decides to cut traffic, how credentials are rotated, which fallback model or manual process keeps work moving, and who drafts notifications to clients and the regulator.
| Step | Question to answer | Owner |
|---|---|---|
| Map the chain | Who touches our data, and where? | Procurement with IT |
| Trace data flows | Which features store prompts or outputs? | IT security |
| Contract down | Do sub-processors carry our terms? | Legal |
| Minimise input | What should never enter the tool? | Data owner per function |
| Rehearse | Can we cut traffic and recover in a day? | CIO or COO |
What does the PDPO require when an AI vendor uses sub-processors?
Under the PDPO, a data user that engages a data processor must use contractual or other means to stop personal data being kept longer than necessary and to protect it from unauthorised access. That duty stays with you even when your processor hands the data on to another company.
These obligations sit in Data Protection Principles 2(3) and 4(2) of the Personal Data (Privacy) Ordinance. In practice, the regulator will ask what you did, not what your vendor promised. A contract that never mentions sub-processors, retention periods or deletion is hard to defend after an incident.
The Privacy Commissioner for Personal Data has also published AI-specific guidance, including its model framework and guidance on agentic AI, which expects organisations to assess risk across the AI supply chain. For banks, the HKMA expects outsourcing and third-party risk controls to extend to AI services. Cross-border transfers deserve particular attention when an inference provider runs models outside Hong Kong.
A practical test for your legal team: for each AI tool, can you name the country where prompts are processed, the retention period for each feature, and the clause that binds every sub-processor? If any answer is unknown, Step 1 and Step 3 come first.
How does this framework work in a Hong Kong enterprise?
Start with the AI tools that process the most personal or client data, usually customer service, HR and document review. A mid-sized firm can map and score its top five AI tools in four to six weeks with procurement, IT security and legal working from one shared register.
Consider a Hong Kong wealth management firm with 300 staff. It uses an AI meeting assistant for client calls, an AI writing tool inside its productivity suite and a chatbot on its website. Mapping the chain shows that the meeting assistant sends audio to a transcription provider in another jurisdiction and stores summaries in an analytics feature outside its zero data retention terms.
The fix is not to ban the tool. The firm switches off the analytics feature, adds a flow-down clause at renewal, restricts the assistant to meetings without identity documents on screen, and names the COO as owner of the vendor-breach runbook. Total cost: a few weeks of staff time and one contract amendment.
A logistics group faces a different chain. Its AI agents connect to shipment systems and email, so the main risk is access tokens rather than transcripts. Here Step 4 matters most: short-lived credentials, scoped permissions and a tested process to revoke every token the AI vendor holds.
What mistakes do enterprises make with AI vendor due diligence?
The most common mistakes are treating zero data retention as a blanket guarantee, reviewing vendors once at signing and never again, accepting a sub-processor list without checking what each party does, and having no plan to cut off a vendor quickly when its own supplier is breached.
- Trusting the headline term. Zero data retention on the main API says nothing about dashboards, logs or support tickets.
- Annual reviews in a monthly market. AI vendors change model providers often. Ask for notice of changes, not a yearly questionnaire.
- Security questionnaires without evidence. Ask for the sub-processor list, the data flow diagram and the latest independent audit report.
- No exit route. If you cannot switch off a feature or a vendor within a day, you cannot contain a breach that starts two layers away.
- Shadow tools outside the register. Teams adopt AI features inside existing software without procurement ever seeing a new sub-processor.
How should you report AI supply chain risk to the board?
Report it as a short register: each AI tool, the data it handles, the number of sub-processors behind it, whether retention is confirmed feature by feature, and whether a tested exit plan exists. Show the trend each quarter so the board sees risk falling as adoption grows.
Boards do not need the technical chain. They need to know three things: where the organisation's most sensitive data goes when staff use AI, whether contracts cover every party that touches it, and how fast the organisation can respond when one of those parties fails.
Gartner frames this as balancing value and risk rather than choosing between them. Controls such as access limits, audit trails and deletion checks slow the wrong things down and let the right projects scale. A one-page register makes that trade-off visible and gives the board a reason to keep funding adoption.
Govern the chain, not just the contract
Every AI tool your teams use rests on companies you will never meet. That is not a reason to slow adoption. It is a reason to know who those companies are, what they keep and how quickly you can walk away.
Map the chain, check retention feature by feature, push your terms down, keep sensitive data out and rehearse the bad day. Enterprises that do this now will be able to say yes to new AI tools faster, because they already know which questions to ask.
Empowering AI. Centered on you.
Deploy AI with a data chain you can explain
If you want AI that works for your teams with data kept in Hong Kong and a chain your compliance team can sign off, start with a free AI Staff configuration assessment. We'll walk you through every step, from mapping your first workflow to deployment and day-to-day operation.
Book a Free AI Staff Assessment
Frequently Asked Questions
Does zero data retention mean my data is never stored?
No. Zero data retention usually covers specific endpoints or products. Features such as admin dashboards, history, fine-tuning or support logs may store data under separate terms, so confirm retention for each feature you use.
Should we stop using AI tools that rely on third-party inference providers?
Not necessarily. Almost every AI tool relies on other providers. The goal is visibility and control: a known sub-processor list, flow-down contract terms and the ability to cut traffic quickly.
How often should we review an AI vendor's sub-processor list?
Ask for notice whenever the list changes and review it at least quarterly for tools handling personal or confidential data. AI vendors change model providers far more often than traditional SaaS vendors.
Who should own AI sub-processor risk inside the organisation?
Give one executive, usually the CIO or COO, overall ownership, with procurement, IT security and legal each owning their step. Shared ownership without a single name tends to mean nobody acts during an incident.