01
We receive the original text
Your application sends the request to QDivZero. With Paranoia Mode enabled, we check it before sending it to the provider running the model.
PARANOIA MODE / OPTIONAL PROTECTION
Mask sensitive data.Before inference.
Paranoia Mode is an optional protection you can enable. When enabled, QDivZero checks your request text and replaces detected passwords and sensitive personal data with placeholders. The model receives that version of the request, without the masked values. This protection complements zero logs and encrypted communications: our provider does not store inference data, and this mode also reduces the sensitive information it needs to process.
01 / HOW IT WORKS
How we protect your request when you enable it.
If you choose to enable Paranoia Mode, the check takes place on QDivZero servers. We look for sensitive data in the text and replace detected values with placeholders. The inference provider, which runs the model, receives the text with those replacements.
02
We replace detected data
We replace detected passwords and sensitive personal data with placeholders. Instructions and the rest of the text remain in the request.
03
We send the protected version
The provider runs the model on the modified text and generates the response. Values we have masked are not included in that request.
When enabled, Paranoia Mode reduces the sensitive information sent to the provider. Zero logs avoids logging inference content at QDivZero by default, and our provider does not store it. HTTPS/TLS protects communications in both directions. All three measures complement each other; zero logs and encryption remain in place with Paranoia Mode disabled.
02 / AN EXAMPLE OF THE DATA PATH
The request changes before it leaves.
In this example, the model can work with the message without knowing the original name, email or password. Compare both versions: the provider does not store inference data in either case. All example data is fictional.
QDivZeroMasking applied at QDivZero
Local simulation with fictional data. The switch only changes this example. Labels explain the masking concept. Actual results depend on detected data and the available configuration.
What gets filtered
If you enable Paranoia Mode, we analyze supported request text on our servers and mask detected sensitive data before contacting the provider. The rest of the context still reaches the model.
What to review in your integration
Check which request parts and data types the available configuration supports. Do not assume that every kind of content, format or file is protected by this detection.
When a name is replaced by a placeholder, the model receives that placeholder and cannot learn the original name from that text. Pre-inference filtering does not change storage settings or delete data already saved.
03 / DEFENCE IN DEPTH
Zero logs and one more layer.
Zero logs, Paranoia Mode and HTTPS/TLS complement each other to protect your data: they control what is retained, what information reaches the model and how it travels over the network. Our provider does not store inference data. You can enable Paranoia Mode as an optional layer to also reduce the sensitive information it processes.
RETENTION
Zero logs
QDivZero does not log prompts or responses by default, and our provider does not store inference data. The exception at QDivZero is the Responses API, which retains context to continue conversations.
OUTBOUND CONTENT
Paranoia Mode
If enabled, we mask detected sensitive data before sending the prompt to the provider. The remaining context still travels to the destination.
COMMUNICATIONS
HTTPS/TLS encryption
Encrypts requests and responses on both connections: application–QDivZero and QDivZero–provider. Protects against interception in transit. At the destination, the model processes the version of the prompt we send.
What does it add if the provider does not store data?
To generate a response, the infrastructure needs to process the request even though it does not retain it afterwards. If you enable Paranoia Mode, we reduce the sensitive information present during processing. In the event of hypothetical unauthorised access while the request runs, masked values would not be part of the content sent to the provider. This adds to zero logs and encryption in transit; its scope depends on which data is detected and masked.
Explore all security measures04 / FAQ
What Paranoia Mode protects.
Masking, storage and encryption protect different parts of your data’s journey.
What does it add to zero logs?
They work together. Zero logs avoids logging inference content at QDivZero by default, and our provider does not store this data. If enabled, Paranoia Mode acts earlier: it masks detected sensitive data to reduce the information that reaches the model. The storage policy remains in place whether this layer is enabled or disabled.
If the provider does not store data, why add this protection?
Because it needs to process the request content to respond. Not retaining it afterwards reduces storage risks; masking data before sending it reduces exposure during execution. For example, a model can summarise an issue using [PERSON] and [EMAIL] without receiving the original name or email address.
How does masking data differ from encrypting the connection?
HTTPS/TLS protects requests and responses while they travel over the network. At the destination, content is decrypted for processing. If you enable Paranoia Mode, we replace detected sensitive data before that transfer: the model receives the protected version and masked values are not part of that request. Both protections work together.
Who receives the original prompt?
QDivZero receives the original request. If you enable Paranoia Mode, we perform detection on our servers and the provider receives a version with detected elements masked. The rest of the prompt remains part of processing.
What if the provider has a security incident?
The provider does not store inference data. If unauthorised access occurred during processing, values masked by Paranoia Mode would not be part of that request. This reduces their exposure; protection depends on which information was detected and masked.
Does it detect every confidential detail?
Perfect detection should not be assumed. Passwords and personal data can appear in different formats and contexts. Review the scope available for your integration and avoid sending secrets your application does not need to process.
Is it enabled automatically?
You choose whether to enable it. Masking only applies when Paranoia Mode is enabled. Check the platform controls available for your integration. The example on this page explains the feature and does not activate any account configuration.
Is this the same as anonymisation or GDPR compliance?
No. Masking detected elements reduces shared information, but remaining context may still identify a person. It is an additional technical measure that supports data minimisation and needs to be assessed alongside your application’s other requirements.
What about context retained by Responses?
Masking outbound content and retaining context are separate controls. Review Responses storage options. You can delete history or disable its storage globally and use Chat Completions.
Share what is needed. Protect what is sensitive.
You can enable Paranoia Mode to add protection before sending requests to the inference provider. You decide whether to use this extra layer.
Access the platform