platform/security-governance
The question is not whether it is encrypted. It is what crosses the boundary.
Everyone encrypts at rest. The questions a security review actually turns on are narrower: which fields leave your tenant, what a model provider retains, who can approve a regulated claim, and what any of that looks like six months later.
Only the fields an agent's task requires cross to a provider · nothing is retained on the far side
Contractual, not a setting you have to find.
No shared corpus, including for benchmarks.
Which human, which policy, which rule-set version.
At rest. TLS 1.2 or better in transit, internally too.
Four lenses a security review actually uses.
Grouped the way your reviewer will ask, rather than the way a marketing page usually organises them.
What is sent to a model, and what is not.
No training on your data
Your catalog, documents and copy are never used to train or fine-tune a model, ours or a provider's.
Zero-retention inference
Model providers are used under zero-retention terms, so prompts and completions are not stored on their side.
Minimised payloads
Only the fields an agent needs for its task cross the boundary. A copywriting call does not carry your cost price.
Deterministic gates
Model output passes strict schema validation before it becomes a value. Non-conforming output is rejected, not repaired.
Grounded generation
Agents write from resolved attributes with sources. A claim without an attribute behind it does not publish.
Provider portability
The model provider is an implementation detail behind the agent interface and can be changed without touching your data.
Payload minimisation does more for your risk posture than any encryption statement. A copywriting call carries the attributes needed to write copy — it does not carry your cost price, your margin, your supplier terms or anything identifying a person. The safest data is the data that was never sent.
Six things this platform will never do.
Stated as prohibitions rather than features, because that is the form your legal team needs them in.
Train any model on your catalog, documents, imagery or copy.
Pool your data with another customer's for any purpose, including benchmarks.
Send a field to a model that the agent's task does not require.
Publish a value that cannot be traced to a source document.
Write to a system or channel outside the scopes you explicitly enabled.
Hold your enriched data hostage. Full export is available at any time, including on the way out.
Each of these appears in the contract, not only on this page. A commitment that exists exclusively in marketing copy is not a control, and your reviewer will say so.
The last one is the one we would encourage you to test hardest, with us and with every other vendor you evaluate. Ask precisely what leaving looks like, and how long it takes to get your enriched attributes and their provenance out.
Send us your security questionnaire.
We would rather answer your standard vendor assessment than have you read ours. Send it as-is and we will return it completed, with the SOC 2 Type II report and a data flow diagram for your specific integration, under NDA.
Your questionnaire · SOC 2 Type II under NDA · Response in 5 working days