# Zero Data Retention Is Not Zero Data Exposure

> Source: <https://tinfoil.sh/blog/2026-11-16-zero-data-retention-is-not-zero-data-exposure>
> Published: 2026-11-16 12:00:00+00:00

# Zero Data Retention Is Not Zero Data Exposure

Attackers [breached Fireworks AI’s systems](https://fireworks.ai/security/bulletins/october-2026-security-incident-update) and extracted customer data. Good thing Fireworks advertises [zero data retention (ZDR) by default](https://docs.fireworks.ai/guides/security_compliance/data_handling), right?

Shortly after the hack, Notion sent out a [recommendation](https://www.notion.com/security/advisories/security-incident-involving-a-third-party-ai-vendor) to rotate potentially exposed passwords, API keys, and other credentials. Notion later disclosed that among the data exposed from Fireworks’ servers were [Notion AI chat transcripts](https://www.notion.com/security/advisories/security-incident-involving-a-third-party-ai-vendor), which, according to Notion, Fireworks had [contractually committed](https://www.notion.com/help/security-incident-involving-a-third-party-ai-vendor) not to retain.

So what happened then? Notion’s AI transcripts were sent to Fireworks for classification jobs behind Insights (a storage-dependent feature outside of [Notion’s enterprise ZDR commitment to its own customers](https://www.notion.com/help/security-incident-involving-a-third-party-ai-vendor)). As it turns out, ZDR is not an airtight guarantee. For example, [Fireworks’ default privacy policy](https://docs.fireworks.ai/guides/security_compliance/data_handling) has several exceptions around ZDR, for things like batch inference and stored Responses. Fireworks CEO Lin Qiao said that following the hack, they found no impact to [“Inference (excluding batch inference).”](https://x.com/lqiao/status/2119943330612256890)

Such parenthetical exceptions and exclusions make it difficult to know when, how, and where a ZDR guarantee applies to the AI services you use. And when stacked together across vendors, they result in a Swiss cheese model of security.

Swiss cheese security is bad. For enterprise customers, ZDR is now a prerequisite for the approval of AI services. Aaron Levie (CEO of Box) recently posted about this:

ZDR lets providers claim to customers that their data is safe. But unfortunately, as a technical guarantee, ZDR is very weak. It does not restrict what a provider can read, learn, or keep. A contract rules out those actions on paper, but the provider’s systems still have to enforce those restrictions.

## Contracts don’t compose well

Requests rarely touch only one company. Instead, each router, model host, and tool that a request flows through can have its own retention policy. If you’re lucky, the request holding your data will jump through the right ZDR hoops and get purged after completion. But if you’re unlucky, well, ZDR is as good as saging your data to ward off evil bits.

Notion’s customers were two hops away from Fireworks, where [their transcripts were exposed](https://www.notion.com/security/advisories/security-incident-involving-a-third-party-ai-vendor). The attackers entered through a fourth company: the [cloud provider hosting Fireworks’ R&D machines](https://fireworks.ai/security/bulletins/october-2026-security-incident-update). Each provider handles these boundaries differently. [Together AI disables passthrough models](https://docs.together.ai/docs/zero-data-retention) when ZDR is enabled, while [OpenRouter’s ZDR filter](https://openrouter.ai/docs/guides/features/zdr) applies to inference routing and leaves plugins and tools to their own policies. OpenAI warns that data sent to [remote MCP servers](https://developers.openai.com/api/docs/guides/your-data) follows those operators’ retention rules. Confused yet? The deeper you dig into what ZDR stands for, the more carve-outs you find.

## Every provider defines its own zero

Each of those companies offers its own version of ZDR, making it difficult to compare between providers. Despite the obvious definition, in reality, “zero,” “data,” and “retention” can be defined in their own way by providers. This becomes a challenge for developers, since AI applications often cycle between multiple AI providers and services. Even if all of them claim ZDR, the reality is murky.

We read the ZDR documentation and terms of 12 AI providers and adjacent services. Across these policies, every provider sets its own scope, defaults, and eligibility rules, with exclusions that can include sensitive data.

Once you have ZDR, it may or may not cover all endpoints. Careful to not accidentally call a non-ZDR-enabled endpoint! Defaults vary too. For example, [Fireworks’ Responses API](https://docs.fireworks.ai/guides/security_compliance/data_handling) and [Google’s Interactions API](https://docs.cloud.google.com/gemini-enterprise-agent-platform/resources/zero-data-retention) both store conversation state unless you remember to pass `store=false`. Some providers also **fail open** and let requests proceed even when the endpoint falls outside the ZDR guarantee. [Anthropic doesn’t block ineligible features](https://platform.claude.com/docs/en/manage-claude/api-and-data-retention) for ZDR organizations, so a call to code execution or MCP connectors doesn’t fall under ZDR.

Some providers like [Google](https://ai.google.dev/gemini-api/docs/zdr) and [OpenAI](https://developers.openai.com/api/docs/guides/your-data) also allow prompt caching (with data still accessible) for up to 24 hours under ZDR, and [OpenRouter](https://openrouter.ai/docs/guides/features/zdr) does not count in-memory caching as retention at all. OpenRouter’s ZDR enforcement can also route to Safety Retention endpoints, where, [per OpenRouter’s own docs](https://openrouter.ai/docs/guides/features/zdr), the provider may retain prompts and responses from any request for abuse prevention. Self-serve accounts allow those endpoints by default unless the customer explicitly selects “ZDR only.” Even then, can you be sure that all the leaks are plugged?

Many are surprised to learn that just because data isn’t retained doesn’t mean it can’t be used for training. Some providers are more honest about this than others. For example, Mistral is explicit about separating “no training” from “ZDR,” labeling them as [separate controls](https://docs.mistral.ai/admin/monitor-comply/zero-data-retention). Its [commercial terms](https://legal.mistral.ai/terms/commercial-terms-of-service) let it train on some model traffic even when a customer has enabled ZDR elsewhere.

| Provider | What ZDR covers | Limits | 
|---|---|---|
| [OpenAI](https://developers.openai.com/api/docs/guides/your-data) | No content in abuse logs; Responses and Chat Completions aren’t stored. | Stateful features, a 24-hour prompt cache, and flagged CSAM images. Some frontier models require [Private Safety Processing](https://developers.openai.com/api/docs/guides/private-safety-processing) , which keeps flagged or sampled records in your storage for at least 30 days. | 
| [Anthropic](https://platform.claude.com/docs/en/manage-claude/api-and-data-retention) | Eligible prompts and responses discarded after the reply. | Ineligible features still run, outside ZDR. Some frontier models keep data for 30 days unless approved; flagged content up to 2 years. | 
| [Google](https://docs.cloud.google.com/gemini-enterprise-agent-platform/resources/zero-data-retention) Agent Platform | Varies by agreement and model; guaranteed only on Agent Platform, not the [Gemini API](https://ai.google.dev/gemini-api/docs/zdr) . | Search and Maps grounding kept 3 to 30 days; research sessions 7 days; prompt cache stored at rest by default. | 
| [Mistral](https://docs.mistral.ai/admin/monitor-comply/zero-data-retention) | Supported stateless API calls. | Stateful features excluded; Labs and Preview traffic may be [used for training](https://legal.mistral.ai/terms/commercial-terms-of-service) . | 
| [xAI](https://docs.x.ai/developers/faq/security) | Inputs and outputs never written to disk; stateful features disabled. | Off by default; requests are otherwise kept for 30 days. | 
| [Microsoft](https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/abuse-monitoring) Foundry | Abuse-log storage and human review turned off for approved customers. | Automated review continues; [feature storage](https://learn.microsoft.com/en-us/azure/foundry/responsible-ai/openai/data-privacy) still applies. | 
| [AWS](https://docs.aws.amazon.com/bedrock/latest/userguide/data-retention.html) Bedrock | No durable storage of requests or responses. | Models that require retention are blocked; required human review stays inside AWS. | 
| [Together AI](https://docs.together.ai/docs/zero-data-retention) | Request content kept out of durable storage. | Off by default and not retroactive; uploaded batch and training files stay until deleted. | 
| [Fireworks AI](https://docs.fireworks.ai/guides/security_compliance/data_handling) | Open-model inference held in memory, by default. | Batch jobs and [dataset-based training](https://docs.fireworks.ai/guides/security_compliance/secure_training) use storage; Responses kept up to 30 days;[enforced policy](https://docs.fireworks.ai/accounts/zero-data-retention) requires Enterprise. | 
| [OpenRouter](https://openrouter.ai/docs/guides/features/zdr) | Routes only to ZDR endpoints when selected. | Self-serve defaults allow Safety Retention endpoints; tools excluded; own-key declarations unverified. | 
| [Exa](https://exa.ai/docs/admin/security/zero-data-retention) | Search, Contents, Agent, and Answer, on Enterprise. | Websets excluded; [default privacy policy](https://exa.ai/privacy) allows training on queries. | 
| [Firecrawl](https://docs.firecrawl.dev/security) | Content held only for the request or session, on Enterprise. | Anonymized search lets the upstream provider cache queries. | 

## “Just trust us”

It should be clear by now that zero retention is not synonymous with privacy. Moreover, even if every provider kept its word and diligently deleted each request upon completion, you would have no way to know. Policies can change, and even though providers like OpenRouter promise [advance email notice](https://openrouter.ai/privacy) of material changes affecting registered users, are you ready to monitor all endpoints and make sure that the latest policies match what you expect?

[Fireworks’ statement](https://fireworks.ai/security/bulletins/october-2026-security-incident-update) that its ZDR products were untouched is just another self-report. No customer can confirm it, and implementation bugs can void it. None of the policies we read offers a way to prove that nothing was kept. OpenAI recently announced its [Private Intelligence](https://openai.com/business/private-intelligence/) offering to select enterprise customers. While it claims verifiable controls through confidential computing technology, it appears that [parts of their system still rely on ZDR](https://openaiassets.blob.core.windows.net/$web/pdf/c7284810-2252-462f-803e-075b0c95bccb/psp-whitepaper.pdf#page=10) for safety processing.

## Building end-to-end verifiable privacy

ZDR promises not to keep your data. Tinfoil makes it impossible to see it in the first place. Inference runs in secure hardware enclaves with confidential computing VMs. Your data stays inaccessible even while it’s being processed, so the provider can’t see it. [Remote attestation](https://tinfoil.sh/security-and-privacy-faq) proves exactly what code is running during the entire inference pipeline, and because the [security-critical code is open source](https://tinfoil.sh/blog/2026-09-14-safety-without-compromising-privacy), customers can inspect the software and [verify the processes handling their data](https://tinfoil.sh/blog/2026-10-25-less-is-more). These protections come from the code and hardware implementing policies that don’t need to be trusted.

Fireworks’ customers found out what “zero” meant after the breach. We built Tinfoil so you can check before you send a single prompt.

### Subscribe for Updates

[RSS Feed](https://tinfoil.sh/feed.xml)

Stay up to date with our latest blog posts and announcements.
