AI DLP: keep sensitive data out of external models
Classic DLP watches email and file shares. AI DLP watches prompts, attachments and model responses, and can mask rather than block, so work continues while the sensitive values never leave your perimeter.
People paste what they have in front of them: a customer email with a card number, a log with an access key, a contract marked confidential. For a model to summarise or rewrite that text, it rarely needs the real values.
Rexaris AI detects sensitive entities on the way out, replaces them with stable placeholders such as <PERSON_1> or <CARD_1>, and restores the originals inside your perimeter when the response comes back. Where masking is not acceptable, the request is blocked with a clear reason.
Why AI traffic needs its own DLP layer
Prompts are unstructured
A card number or a key can appear anywhere in free text, code or an attached document. Detectors built for AI traffic look at the whole request body, not a file type.
Mask instead of block
Most prompts work just as well with placeholders. Masking keeps adoption high; blocking is reserved for data that must never leave in any form, such as credentials.
Responses leak too
Models can echo secrets back or generate them from context. The response path is inspected with the same detectors.
Custom patterns
Account formats, internal identifiers and project codenames are added as your own classes with your own actions.
How AI DLP works in Rexaris AI
Detection
PII, PCI data, API keys, credentials, secrets, IP addresses, source code and custom patterns.
Learn moreMasking with restoration
Placeholders are consistent within a conversation and restored inside the trusted boundary.
Learn moreRouting as an alternative
Confidential material can be sent to an internal model instead of being blocked.
Learn moreEvidence
Each masked or blocked entity is counted in the audit log and the dashboard.
Learn more
Frequently asked questions
- Does masking break the model's answer?
- Rarely. Placeholders keep the structure of the text, and the same entity always maps to the same placeholder in a conversation, so the model can refer to <PERSON_1> consistently. The user sees real values after restoration.
- Which data classes are detected out of the box?
- Personal data, payment card data, API keys and tokens in common formats, credentials, private keys, IP addresses and source code, plus any regex or dictionary classes you define.
- Is this PCI DSS or GDPR compliance?
- Rexaris AI helps enforce the controls those regimes expect for AI traffic, such as keeping card data away from third parties. It is not a certification and does not make claims about your overall compliance.
- Can it see data sent to a personal ChatGPT account?
- Not by itself; it controls traffic that passes through its gateway. It can ingest shadow AI events from your endpoint, DLP and SASE tools to show that usage alongside managed traffic.
Related solutions
From the blog
2 min read
Why enterprise AI control starts at the gateway
Most companies already have AI inside the building. The question is not whether to allow it, but where the single point of control should sit.
Read2 min read
Mask, block or route: choosing the right action for sensitive data
Three actions cover almost every data policy for external LLMs. Picking the wrong one either leaks data or kills adoption.
Read
Find out what your prompts actually contain
A week of audit-only classification on one team usually settles the policy debate with numbers.