If it's sensitive, we spot it before it reaches the AI.
Tracehold’s detection engine runs in the browser and checks what people type, paste and upload to AI tools before it is sent, with more than 1,600 rules for credentials, identity numbers, bank and card details, health data and confidentiality markings, plus your own keywords.
The sensitive data your team shares without thinking
These are the everyday leaks, pasted into a chatbot to save five minutes. Tracehold knows what each of them looks like.
Passwords & keys
Login credentials, API keys and access tokens for cloud, code and AI accounts: the kind of secret that should never be pasted anywhere.
Personal & health data
Identity numbers, dates of birth and email addresses, in a recognisable form, plus health data such as patient numbers and diagnosis codes.
Financial details
IBANs and payment cards, validated before they count.
Confidential markings
Confidentiality markings and deal material, such as STRICTLY CONFIDENTIAL / TOP SECRET and M&A leads under an NDA.
Not everything has a fixed format
A card number always looks like a card number. A password in a sentence, or a health condition next to a person's email address, does not. Some rules read the context around a value instead of only its shape.
On the Governance plan (beta), an optional LLM classifier can be switched on for your organisation, on top of the rules. When it is on, the text travels in transit to the classifier, and Tracehold does not store it.
Secrets stated in plain words
A rule looks for a password or secret written into a sentence, even when no key name or format gives it away.
Values with no known shape
A statistical check (entropy) flags strings that look like a secret even when no rule knows the format. It is a low-severity signal on purpose.
Context decides
GDPR special categories, such as health, religion or union membership, are flagged only when a personal identifier is nearby. The same words on their own are not a finding.
Your own keywords
Project code names, client codes and internal product names are added in the console under Prevention, Detection library, Sensitive entities. Your policy decides what happens when one shows up.
Flagged: special category near personal data
A health condition next to a personal identifier is a finding (gdpr_art9_with_pii).
Left alone: the same words on their own
“He has type 2 diabetes.” has no personal identifier nearby, so it is not flagged.
A password in a sentence
“The staging server password is Zq8!mVx2pL.” is flagged even with no key name or format.
It reads screenshots and files, not just typed text
A secret hidden in a screenshot or a PDF is still a leak. The extension reads the text inside images and documents before they are uploaded and checks it against the same rules.
Screenshots & images
Text in an image you paste or upload is read with OCR in the browser and checked like any other text.
Documents & attachments
PDF files and current-format Word, Excel and PowerPoint files (not the older .doc, .xls and .ppt) are opened in the browser and their text is checked. A scanned PDF with no text layer goes through OCR.
Same rules, same policy
What is read from a file is checked against the same rules as typed text, and the result follows your policy.
Real product recording, illustrative sample data. A .env file is caught and the upload is blocked before it reaches the AI tool.
Alerts you can trust, not noise you'll ignore
The fastest way to make a security tool useless is to flood people with false alarms. The rules are built to be precise, and you can tune the rest.
It checks, it doesn't guess
Identity, IBAN and card rules (more than 30 rules in all) validate the check digit, not just the shape of the number. A valid PPS number is flagged; the same number with a wrong check letter is not.
Examples and placeholders are left alone
Known example values and placeholder words are recognised and ignored, so documentation snippets and test data do not raise alerts.
You can tune it
Choose how strict the policy is (soft, moderate or aggressive), switch rules on or off, change a rule's severity and add exceptions for the cases that matter to you. All from the console.
Don't just block it. Hide the sensitive part and let the rest go
Blocking everything frustrates people and pushes them to tools you cannot see. When your policy says Redact, Tracehold replaces only the sensitive value with a marker and lets the rest of the prompt carry on.
The employee keeps working, and the AI tool can still answer the actual question. Before a redacted prompt goes out, Tracehold checks that no original value is left in it; if one is, the send is stopped instead.
The rest still works
Only the value is swapped out, so the AI tool can still answer the actual question.
A final check
Before a redacted prompt goes out, Tracehold checks that no original value is left in it. If one is, the send is stopped instead.
A record, not the prompt
The console receives event metadata plus a masked sample of the value. Keeping the content is off by default.
The same protection, wherever AI gets used
The browser extension is the core. The agents extend it to the places a browser cannot see, and everything reports to one console.
The browser
Chrome, Edge and Firefox. The extension recognises more than 300 AI tool domains, including ChatGPT, Claude, Gemini and Copilot. A name-based check also helps catch new AI tools that are not in the catalogue yet.
The desktop
The Windows agent covers USB drives, printing, screen captures of protected Office documents and on-demand file scans.
Developer tools
The Linux agent covers terminal AI tools such as Claude Code and Codex, when your policy turns it on.
What buyers ask about detection
What kinds of sensitive data does it actually catch?
What about data that doesn’t follow a fixed pattern?
Will it flood my team with false alarms?
Does it scan images and documents, or only typed text?
If something sensitive is found, do you block the whole message?
Where does the detection run, and what reaches Tracehold?
See what Tracehold would catch in your environment
Request a demo and we’ll show you, on your own data, what would have been spotted. Start with the Observe action before you enforce anything.