AI security: What Biscuit’s AI can see and keep

Biscuit’s AI builds and runs your apps, so it’s designed to keep your data and credentials safe while it works. This page explains which AI models Biscuit uses, what they can and can’t see, and how your data is handled by Biscuit and its AI providers.

​

How Biscuit uses AI

Biscuit uses AI in two places:

• Building your apps: when you describe what you want in chat, Biscuit’s AI writes the code behind your app: its screens, logic and data models.
• Inside your apps: your apps can use AI as a step in a workflow, for example to write a summary, generate an image or create a voiceover.

​

Which AI models Biscuit uses

Biscuit uses leading models from OpenAI, Anthropic and Google, such as GPT-6, Claude Opus 5.5 and Gemini 4, plus a range of open-weight models. Biscuit tests models on each kind of task and uses the best one for every step. New models are added as they’re released.

​

How your data is protected

• While building your app: the AI can read the results of queries it runs while it builds, for example to check that a feature works.
• Between workspaces: each workspace has its own isolated database. The AI working on your apps can only reach your workspace’s data.
• Your API keys and client secrets: the AI never sees your keys, and they’re never sent to AI model providers or to models Biscuit hosts. When your app calls an outside service, the request goes through Biscuit’s proxy, which adds the key at that moment and only allows calls to approved domains. A key can’t be sent anywhere else.
• Data retention by AI providers: most models Biscuit uses have zero data retention (ZDR). The provider processes your data only to generate a response and keeps nothing afterward. A small number of models don’t offer zero data retention. When Biscuit uses one of these, the provider may keep your data for a limited period, only to deliver the response. Opting in for ZDR for all providers that we use is on the roadmap.
• Data residency: Biscuit’s infrastructure currently includes servers in the US. EU-only hosting is on the roadmap. See Security.

​

Does Biscuit use my data to train AI models?

To get better at building apps, Biscuit may use data from your use of the product to train and improve its own AI models. Before any data is used, personal information is removed or anonymized, and Biscuit never trains on personal data. This also means that no company data is included in any training. This applies only to Biscuit’s own models. AI providers such as OpenAI, Anthropic and Google can’t use your data to train theirs. Biscuit’s contracts with AI providers don’t allow them to use your data to train or evaluate their models. This covers every provider involved in a request: the company that made the model and the company that hosts or serves it.

​

Opting out

You can opt your workspace out of training at any time.

1. Email support@biscuit.so and ask to opt your workspace out.
2. Our team turns training off for your workspace.
3. Check your status in your workspace settings.

​

How AI changes are reviewed

• Data access rules are checked: when the AI changes how an app stores data, a separate security review checks for exposed personal data and for permissions that are broader than they should be.
• Nothing goes live without you: AI changes create a new version that only builders see. People using the app keep the published version until someone with Can edit access publishes again.
• A safety check before going live: when you publish an app, Biscuit runs an automated safety check on its configuration before it goes live.

​

Compliance

For hosting location, SOC 2, ISO 27001 and penetration testing, see Security and the Trust Center. For more on the topics that matter to IT and procurement teams, see IT and procurement FAQ. Have more questions? Email us at support@biscuit.so.

​

FAQ

​

Up next