# Grodmir — Prompt DLP for sensitive AI data

> Prompt DLP means data loss prevention for AI prompts. Grodmir inspects sensitive data in connected AI channels and applies signed policies to remove detected secrets, pseudonymise data or block a request before forwarding.

Grodmir is an AI gateway for data protection and AI usage governance in organisations.

## What does Prompt DLP do?

Prompt DLP means data loss prevention for AI prompts. Grodmir inspects sensitive data in connected AI channels and applies signed policies to remove detected secrets, pseudonymise data or block a request before forwarding.

[Answer source](https://grodmir.com/en/prompt-dlp#what-does-prompt-dlp-do)

## Are secrets ever restored?

Detected passwords, API keys, private keys and tokens are removed irreversibly in enforce mode. They never enter pseudonym maps and are never restored; observe mode forwards unchanged content.

[Answer source](https://grodmir.com/en/prompt-dlp#are-secrets-ever-restored)

## Which languages and identifiers are supported?

Detection supports Polish, German and English. Checksum rules cover identifiers including PESEL, NIP, REGON and IBAN; names, organisations and addresses use a local model trained on synthetic data. Pseudonymisation is reversible under permitted response restoration and is not anonymisation.

[Answer source](https://grodmir.com/en/prompt-dlp#which-languages-and-identifiers-are-supported)

## When should a team evaluate Grodmir Prompt DLP?

Evaluate Grodmir when your team needs policy-based checks for sensitive text in connected AI requests, with local inspection and irreversible removal of detected secrets in enforce mode. Start in observe mode to measure misses and false alarms; it forwards content unchanged.

[Answer source](https://grodmir.com/en/prompt-dlp#when-should-a-team-evaluate-grodmir-prompt-dlp)

## What would a Prompt DLP transformation look like?

Under a signed policy that pseudonymises detected email addresses, Grodmir can forward “Contact [EMAIL_1]” instead of “Contact alex@example.com”. This synthetic example illustrates policy behaviour, not a measured detection result. In observe mode the original text is forwarded unchanged; detected secrets are never added to a restoration map.

[Answer source](https://grodmir.com/en/prompt-dlp#synthetic-prompt-example)

## How should a team measure Prompt DLP detection?

Label representative samples independently before running the detector, including samples with no sensitive data to measure false alarms. Use aiguard eval locally and review results by category, language and source. The measurement covers labelled detection, not policy decisions, client dictionaries, repository rules or traffic that bypasses the gateway.

[Answer source](https://grodmir.com/en/prompt-dlp#measure-prompt-dlp-detection)

## How does secret removal differ from pseudonymisation?

In enforce mode, detected secrets are removed irreversibly or cause a policy block. They never enter a pseudonym map. Other detected sensitive values can be replaced with placeholders such as [EMAIL_1]; their originals remain in a local, expiring map for permitted restoration in model responses. Pseudonymisation is reversible and is not anonymisation. MCP tool-result placeholders are not restored across later tool calls. Observe mode forwards the original content unchanged.

[Answer source](https://grodmir.com/en/prompt-dlp#removal-and-pseudonymisation)

## Does Prompt DLP cover browser chats and MCP tools?

Coverage depends on the connection. Model API requests must use the Grodmir gateway. The browser extension checks pasted or dropped text on supported chat pages through the customer gateway; typed text and other upload paths remain outside that integration. MCP arguments and results need the MCP proxy or central gateway. Installing one component does not automatically cover the other channels.

[Answer source](https://grodmir.com/en/prompt-dlp#browser-and-tool-coverage)

- [Grodmir — Prompt DLP for ChatGPT in the browser](https://grodmir.com/en/integrations/chatgpt-browser)
- [Grodmir — Claude Code gateway and prompt data protection](https://grodmir.com/en/integrations/claude-code)
- [Grodmir — Codex CLI gateway for the Responses API](https://grodmir.com/en/integrations/codex)
- [Grodmir — MCP proxy for tool data protection](https://grodmir.com/en/mcp-proxy)
- [What do the Grodmir categories mean?](https://grodmir.com/en/glossary)
- [Grodmir — AI gateway selection criteria](https://grodmir.com/en/how-to-choose-an-ai-gateway)

## How can you verify this behaviour?

- [browser-extension (EN)](https://grodmir.com/dokumentacja/zrodla/browser-extension.md)
- [claude-code (EN)](https://grodmir.com/dokumentacja/zrodla/claude-code.md)
- [codex (EN)](https://grodmir.com/dokumentacja/zrodla/codex.md)
- [mcp-proxy (EN)](https://grodmir.com/dokumentacja/zrodla/mcp-proxy.md)

Content updated: 2026-11-14

Protection covers connected-channel traffic only; detection can miss data and raise false alarms. [Grodmir — AI gateway protection limits](https://grodmir.com/en/limitations)

Source: https://grodmir.com/en/prompt-dlp
