# Security: what Peeve never sees

> Read-only by default, the model never holds credentials, every action is confirmed and signed. SOC 2 Q1 2027; BAA not offered today.

Canonical: https://peeve.ai/product/security/

| Data | Stored | Retention |
|---|---|---|
| Element fingerprints, route map | Yes | Until you delete |
| Session step traces | Yes | Your setting · 0–90 days |
| Screenshots (masked selectors removed) | Optional | Your setting |
| Users' passwords or session tokens | Never | — |
| Customer data used for model training | Never | — |

Prompt injection is mitigated by separating instruction from content and letting the model request capabilities only by name; an injection suite runs in CI on every deploy. SOC 2 Type II is in progress (report expected Q1 2027). A signed DPA is available on Scale. A BAA is not offered today.


## More questions

### Is Peeve GDPR compliant?

Peeve publishes its sub-processor list and data map, offers a signed DPA on the Scale plan, and never trains models on your data. SOC 2 Type II is in progress with a report expected Q1 2027, and a BAA is not offered today.

### Does Peeve see my users' passwords or PII?

Never sees passwords or session tokens. The model requests a capability by name and your execution service makes the call with the credentials, so it never holds a secret. Masking selectors remove sensitive fields and screenshots are optional and masked.

### Does Peeve train on my data?

No. Peeve never trains models on your data. It stores element fingerprints and a route map to navigate your product, plus session traces on a retention window you set, and customer data is never used for training.

### Is AI customer support GDPR compliant?

Look for a published sub-processor list, a signed DPA, and no training on your data. Peeve publishes its sub-processors and data map, offers a signed DPA on the Scale plan, and never trains models on your data. Screenshots are optional and off by default, with masked selectors removed before an image leaves the browser.

### Is AI customer support SOC 2 certified?

Ask for the current status rather than a badge. Peeve's SOC 2 Type II is in progress with a report expected Q1 2027, and the security page states exactly what is and is not in place, with nothing claimed early.

### Is AI customer support HIPAA compliant, and can I get a BAA?

Be direct with vendors here. Peeve does not offer a BAA today, so it is not the right fit for a workflow that requires one. It does provide masking, short retention windows and a signed DPA on Scale, and it never touches records it is not given, but a BAA is honestly not available yet.

### Does AI customer support train its models on my data?

It should not. Peeve never trains models on your data and never stores your users' passwords or session tokens. It keeps element fingerprints and a route map to navigate your product, plus session traces on a retention window you set, and customer data is never used for model training.

### Can the AI agent see passwords, payment details, or other PII?

A safe design keeps secrets away from the model entirely. In Peeve the model never sees a password, a token, or your API keys; it requests a capability by name and your execution service makes the call with the credentials. Masking selectors remove sensitive fields, and screenshots are optional and masked.

### How do you prevent prompt injection on a support AI agent?

The structural defense is to separate instructions from content and to scope what the model can request. Peeve treats page text as data, lets the model ask only for named capabilities rather than composing raw requests, and runs an injection test suite in CI on every deploy, so an injection can request a capability but never reach a key or invent an action.

### How do you limit what an AI support agent is allowed to do?

Least privilege plus confirmation. Peeve starts with no ability to act; you grant capabilities one at a time in a config file reviewed in a pull request, writes run inside an allow-list and are confirmed, and a single command disables all acting instantly.

### Is AI customer support safe for a fintech company?

For fintech, the questions that matter are credential handling, confirmation on writes, and auditability. Peeve keeps credentials out of the model, confirms every backend action and marks destructive ones, writes a signed audit entry per action, and offers a signed DPA on Scale. SOC 2 is in progress and a BAA is not offered, so check those against your requirements.