What redaction software works with ChatGPT and Claude?

Redaction software that works with ChatGPT and Claude has to sit in the path of the request, find personal and sensitive data in the prompt or file, and swap it for labelled placeholders before the request reaches the AI tool. NexusNest PromptWall does this for ChatGPT, Claude, Gemini, Copilot, Cursor and Claude Code. One agent on the device routes the traffic through redaction in the NexusNest service, with no browser extension.

What redaction looks like on a real prompt

Here is a finance prompt with an attached spreadsheet, customer_accounts_Q3.xlsx. The employee types it normally. The AI tool only ever sees the version on the right.

Notice what stayed. The balance is not redacted, because the answer needs the amount. The name, the SSN and the account number became [PERSON_1], [SSN_1] and [BANK_ACCOUNT_1]. ChatGPT says plainly that details are redacted and works from IDs and balances, so the reply is still usable.

How redaction with labelled placeholders works

1. The agent covers the AI tool. One lightweight agent on the laptop routes AI traffic for ChatGPT, Claude, Gemini, Copilot, Cursor and Claude Code, plus Codex via ChatGPT sign-in, through redaction before it reaches the AI tool. It is pushed through your MDM.
2. Sensitive spans are found. Personal data such as names, contact details and government IDs, sensitive data such as payment details, health records and credentials, and confidential data such as source code and financials are detected in the prompt.
3. Each span becomes a label. The value is swapped for a placeholder that says what kind of value was there, in flight. Redaction runs in the NexusNest service, in our cloud or self-hosted, and only the redacted text reaches the AI tool.
4. The answer still works. The AI gets the structure of the question and not the values, so it can still sort, summarise, draft and reason.

Labelled placeholders matter. Blanking a value out leaves the AI guessing. A label like [BANK_ACCOUNT_1] tells it what the value is without telling it which account. Each company sets its own policy and can add custom rules for its own terms, and each category can be set to redact, flag, block or off.

Files and images too

Sensitive data does not only travel in typed text. On ChatGPT, Claude and Gemini on the web, PDFs and images are redacted before upload, including names and numbers inside the file, faces and barcodes. Office, text and CSV files are converted to redacted text. See more before and after cases on the examples page.

Claude Code, Cursor and IDE use

Teams often search for Claude Code DLP, and the need is real. Claude Code reads files in a repository and sends them as context, and Cursor sends the open file and often the files around it. That can include .env files, credentials in test fixtures and customer data in seed scripts, none of which the developer typed into a prompt box.

The same device agent covers Claude Code in the terminal and Cursor in the editor, so prompts are redacted before they reach the model provider. Coverage follows the destination the request is going to, not who holds the API key, so a bring-your-own-key setup is redacted the same way. A provider's own privacy mode covers training and retention. It does not redact what was sent, which is why the two are used together. For what each provider says about its own policies, see ChatGPT, Claude, Gemini and Copilot data privacy.

Where this fits with DLP

NexusNest is the control and governance layer for AI, not a DLP tool. It sits next to the DLP you already run and covers the prompt itself. The full comparison is in AI governance vs DLP.

Want to see it on your own prompts? Start with one team and watch what gets redacted. Explore PromptWall →

Frequently asked questions

Related