Clinicians use AI every day. So does your patient data.
Your teams already use AI to draft discharge letters, summarise notes, code referrals and tidy up research data. The risk is what gets pasted into it: patient records, diagnoses, test results, genetic and mental-health data. Tracehold checks what is sent to AI tools and, according to your policy, observes, warns, redacts or blocks. If nothing sensitive is found, the prompt goes through.
The extension checks the prompt and attached files before they are sent to an AI tool.
Event metadata (type of data, severity, destination, action) plus a masked sample of the value. Saving content is off by default.
The extension's engine matches data that has a recognisable shape. Free-form text needs the optional LLM classifier (Governance plan, beta).
Tracehold is built in the EU and EU-hosted.
The fastest way for patient data to leave is a prompt box
Generative AI didn't wait for a clinical safety review to arrive on the ward and in the lab. It came through the browser: helpful, fast, and one paste away from sending special-category health data to a service outside your controls. The intent is almost always innocent. The exposure isn’t.
Patient records & clinical notes
A clinician pastes a case note (name, history, diagnosis, medication) into an AI tool to draft a referral or discharge letter. Each one is special-category health data under GDPR leaving your control.
Test results & research data
Researchers and analysts paste lab results, imaging reports or rows from a study dataset to "explain this finding" or "tidy up this table". Identifiers and outcomes go straight into a prompt.
Code, integrations & documents
Developers paste integration code with embedded credentials into AI assistants to debug an HL7 or FHIR feed. Teams drop in confidential protocols, ethics submissions and incident reports to summarise them.
Built to recognise the data a health organisation can't afford to lose
With more than 1,600 rules running in the browser extension, Tracehold recognises data that has a recognisable shape: record numbers, identity numbers, ICD-10 codes and credentials. Diagnoses and mental-health terms are flagged when a personal identifier is close by.
Special-category health data
Diagnoses and mental-health terms are flagged when a personal identifier, such as an email address or a Spanish DNI, is close by. Genetic variant IDs (rs…) are flagged when genetic context words are nearby. This is the special-category data GDPR singles out for the strongest protection.
Patient identifiers
Identity numbers such as the NHS number, and medical record numbers, ICD-10 codes, prescription numbers and health-card numbers, recognised from their format and the words around them. Check digits are validated, and names on their own are not a finding. A diagnosis in free prose with no identifier nearby needs the optional LLM classifier (Governance plan, beta).
Credentials & confidential docs
API keys and tokens hidden in pasted integration code, plus documents recognised by their markings (STRICTLY CONFIDENTIAL, TOP SECRET, DO NOT DISTRIBUTE) and by your own keywords. The engine does not read a protocol and decide that it is confidential.
Detection runs in the browser, before the send
The extension’s detection engine runs in the browser and checks the prompt and attached files before they are sent. The console receives event metadata plus a masked sample of the value, not the patient record.
No proxy, no network change
The browser extension runs in the browser. The Windows agent does not intercept the network by default, and a clean Linux install does not either.
What reaches the console
Event metadata (type of data, severity, destination, action) plus a masked sample of the value. In the Action column you see Blocked, Warning, Redacted or Logged. Saving the content itself is off by default. Endpoint agents send an excerpt of the text with the sensitive values masked.
The free extension
The free browser extension needs no account and is for one person. It has no console and no policies. Apart from the reports you choose to send, it does not send message content to Tracehold.
Files already on the device, not just what's typed
The Windows agent can run on-demand file scans, and the Linux agent has an optional local scanner, so files already saved on a device (a discharge letter or a research export) can be checked too.
The honest version
Saving content is off by default. If your organisation switches it on, the original text is stored encrypted and shown only through a permission-controlled, audited action; the console shows a version with the detected values masked.
The Security page explains what the console receives and where it runs. See how we protect your data →
Uncontrolled AI use touches the rulebooks that matter most in health
Health data sits at the strictest end of data protection. A single paste can trip more than one obligation at once.
GDPR Article 9: special category
Health data is special-category data under Article 9, with a higher bar for processing and sharing. Pasting it into a public AI tool can be an unlawful transfer to a third party. Checking it before the send helps keep it within your control.
NIS2
Healthcare is named among the essential and important sectors under NIS2, with stronger cyber-risk management and accountability. Demonstrable control over how data leaves to AI tools, backed by an audit trail, is part of showing you manage the risk.
EU AI Act
As clinical and research teams adopt AI, you need to know where and how it's used, and govern it. In the Governance plan, findings map to 8 EU AI Act controls: Art. 26 (AI systems in use by the deployer), Art. 12 (record keeping) and Art. 10 (data governance).
Compliance mapping you can put in front of the board
In the Governance plan, Tracehold maps what it finds to 39 controls across six frameworks (GDPR, the EU AI Act, DORA, NIS2, ISO 27001 and SOC 2) and exports a report for each, in PDF or CSV, with GDPR Article 9 special-category data front and centre for health.
Risk, not theatre
The mapping reflects what your people are actually doing with AI today: what's being caught, how often, how severe, not a point-in-time checklist that's stale by lunchtime.
Evidence on demand
Each finding is kept as a record of its type, severity, destination and action, with a masked sample of the value rather than the full text, so you can show how AI use is controlled without exposing the patient data itself.
Support for your audits, not a certification
The mapping is a product judgement about which detections evidence which obligation. It supports your audits; it is not legal advice or a certification. Review it with your compliance team.
From "we think we're fine" to "here's the picture"
Most health organisations can't answer a simple board question: how much patient data is going to AI right now? Compliance mapping answers it and shows the trend as your controls take effect.
The Governance plan covers GDPR, the EU AI Act, DORA, NIS2, ISO 27001 and SOC 2, with a report for each in PDF or CSV. See compliance in depth →
No proxy, no network surgery.
A hospital or trust can't bolt a new inline proxy into a clinical network on a whim. The browser extension needs no proxy and no network change.
Pushed via your MDM
The console generates the managed-policy files (.reg, .ps1, .json) and the MSI, and there is a guide for Microsoft Intune. Packages and policy come from the console; your MDM or GPO does the rollout.
Start observe-only
Begin in a pilot that watches and reports without blocking anyone. You see exactly what would be caught, and where your real exposure is, with zero disruption to clinical work before you turn on enforcement.
Plugs into your stack
Sign-in with OIDC or LDAP/AD. Findings from the browser extension go to your SIEM over HTTPS (Splunk HEC or CEF), and alerts go to Slack, Microsoft Teams or your own webhook URL.
Start free, scale to governed
Begin with the free extension, add managed visibility and compliance reporting when you're ready.
Insight
Detection, Shadow AI discovery, the console, policies to block, warn or redact, file scanning, your own rules and alerts.
See InsightGovernance
Everything in Insight, plus compliance risk and audit-ready reports mapped to GDPR, NIS2, the EU AI Act, DORA, ISO 27001 and SOC 2, multi-tenant spaces and the LLM detection engine (beta).
Talk to our teamWhat healthcare buyers ask first
Does patient data reach Tracehold?
Do we need a proxy or network changes?
How does this fit GDPR Article 9 for health data?
How does it help with GDPR, NIS2 and the EU AI Act?
Does Tracehold hold any certification?
Does Tracehold map to HIPAA or other health-sector rules?
Where is Tracehold hosted?
Which plan do we need?
See what your organisation is already sending to AI
Start an observe-only pilot and get a clear, honest picture of where patient and research data is leaking to AI tools, with zero disruption to clinical work.