Practical guides on using AI at work while client names, IDs and case numbers stay on your computer, for professional firms, councils and care providers.
Run a privacy proxy for the OpenAI and Claude APIs in Docker: Compose file, first test, base_url, TLS with nginx, logs, updates and a leak check.
Where personal data hides in OpenAI and Claude API requests, three ways to redact it, how to put real values back in streamed replies, and how to test it.
How to watermark a PDF with Word, LibreOffice, Preview or Acrobat: what to write for a rental or bank copy, where to place it, and why it is not redaction.
AI data leakage in LLM apps: how prompts, logs, RAG, agents, injection and provider flaws let personal data out, and the controls that close each route.
Cloudflare AI Gateway DLP can flag or block personal data, not replace it. Nonimo Gateway swaps it for stand-ins and restores the real values in the answer.
Kong AI Gateway handles PII with an Enterprise plugin inside Kong. Nonimo Gateway does it on your own server and puts the real values back in the answer.
A Lakera Guard alternative for personal data: Nonimo Gateway replaces PII before the LLM and restores the real values in the answer, on your own server.
Keep LiteLLM as your router and choose the PII filter behind it: the Presidio guardrail, the built-in options, or the Nonimo Gateway plugin.
LLM Guard's repository is archived. Nonimo Gateway is a maintained proxy that replaces personal data in LLM calls and restores it in the answer.
Presidio gives you the parts to find and replace personal data. Nonimo Gateway is the finished proxy for LLM API calls, real values restored in the answer.
Nightfall DLP blocks or monitors; it does not put data back. Nonimo Gateway replaces personal data in your OpenAI calls and restores it in the answer.
Fourteen PII redaction and data masking tools for LLM apps, grouped by kind: where each runs, what it restores, how it streams and where Nonimo Gateway fits.
Portkey's PII redaction is irreversible by design. Nonimo Gateway replaces personal data in LLM calls and puts the real values back in the answer.
Private AI is now Limina. What its pricing page publishes, how its re-identify API works, and Nonimo Gateway, a proxy you run that puts data back.
A private LLM keeps every prompt in-house, at the cost of hardware and people. When it is worth it, and when pseudonymising API calls is enough.
Protecto prices by API calls and self-hosting is Enterprise. Nonimo Gateway is a proxy on your own server that restores the data in the answer.
Skyflow runs its service for you, buffers the whole answer before streaming and publishes no price. Nonimo Gateway is a proxy you run on your own server.
What is an AI gateway: a proxy that routes, meters and logs your LLM calls. What its logs mean for personal data, and when a private gateway pays off.
How to delete ChatGPT history permanently: one chat or all, then the memories, Library files and shared links it leaves, and the 30 days after.
Turning off Track Changes keeps every edit in the file. To remove track changes in Word permanently, Accept All Changes, delete comments, run Inspect Document.
Yes. ChatGPT files uploaded photos in a Library that deleting the chat doesn't clear. How long ChatGPT, Gemini (18 months), Claude and Copilot keep a photo.
Mask a CSV before AI reads it: drop columns, swap names for random tokens with the key kept apart, generalise dates and postcodes, stop Excel changing IDs.
How to redact in PowerPoint on a copy of the deck: clear speaker notes, hidden slides, comments, master text and chart data, then check the saved file.
How to redact names in Excel before ChatGPT or Copilot sees them: hidden rows and sheets, comments, ID columns, PivotTables and the Document Inspector.
How to redact a PDF so the text is gone, not covered: Acrobat Pro in six steps, Preview on a Mac, and three checks that catch a black box before you send.
How to redact in Word so the words are really gone: replace each mention, clear tracked changes, comments and hidden text, then save a clean .docx.
Remove metadata from PDF and Word files before an AI tool reads them: author, company, template, file paths, photo GPS and XMP, and how to clear each.
What personal data should be redacted? Paste your text and get it back with names, ID numbers, addresses and IBANs covered, ready for AI or to download.
Anonymised, pseudonymised or de-identified? What each term means under the GDPR, HIPAA and Australian law, and what to redact before a file goes to AI.
Often yes, and it turns on where the output is used, not where you sit. Article 2, the deadlines after the Digital Omnibus, and what a deployer owes.
Does ChatGPT share your data? OpenAI says it doesn't sell it, but its policy lists who else receives it, and deleted chats can take 30 days to go.
Yes on Claude Free, Pro, Max and Ultra unless you switch it off in Settings; no on business plans. The 30-day and five-year clocks, and what each keeps.
Google says it doesn't sell your data, but people review some Gemini chats and Google keeps those up to three years, even after you delete them.
Does Copilot train on your data? Work accounts: no, in writing. Personal: the app released August 18, 2026 says no; the older one trains unless you opt out.
Seven insurer and broker forms, opened and read. Two ask about AI, and both want a description. Ten model answers and the evidence, free in Word and PDF.
Free AI acceptable use policy template in Word and PDF: 15 numbered clauses, four annexes, and the whole text on the page. No email, no watermark.
It is a disclosure. Whether you must report it differs in the UK, Ireland, Australia and the US. The tests, the first hour, and the controls that help.
The same questions, answered against the rules that apply where you work.