Client confidentiality, AI prompts and the evidence
Personal data and secrets pasted into a prompt leave the firm the moment it is sent.
You bought Harvey or Legora. Now govern it. EndLayer shows who uses each legal AI tool and on which devices, checks each prompt for personal data and secrets before it is sent, and flags prompt injection in the documents people upload.
Sanction both for the firm, and raise two detectors from Warning to Block.
A firm buys a legal AI assistant after months of review. Then the practical questions arrive: who is using it, what are they typing into it, and what can the firm show a client who asks.
Legal AI security and governance is the control around the tool you chose: seeing its use, checking what goes into it and keeping a record of both. It also covers the AI tools the firm did not choose, which lawyers reach from the same browser.
Personal data and secrets pasted into a prompt leave the firm the moment it is sent.
A document from another party can carry instructions written for the model, not for the reader.
The sanctioned assistant sits beside AI websites nobody reviewed, in the same browser.
Every law firm AI risk review comes down to three things, and law firm AI governance starts there: client data in a prompt, instructions hidden in an uploaded file, and AI tools nobody approved. Lawyers using ChatGPT next to the firm's sanctioned tool is a common case. EndLayer checks each prompt in the browser, flags injected instructions in uploads, and records which AI websites are in use, so your law firm AI security policy applies on every enrolled device.
Write the rule once: warn on, redact or block personal data in a prompt, and add patterns of your own, such as the format of a matter number or a client's name. EndLayer applies it in Chromium browsers and keeps the record.
Harvey and Legora are built into EndLayer as AI websites. Each has its own profile page in the console, with four things on it.
Visits to each regional tenant, with 30 days of prompts beside them, counted per person and per device.
How many prompts to the site were blocked, warned on, redacted and logged, day by day, with the detectors that fired most.
What the prompt injection rule found in files uploaded to the site, listed by verdict.
Sanction, unsanction or block the site from its own page. A block is applied through each browser's managed blocklist.
Prompt checks run in the browser extension, on Chrome, Edge, Brave, Vivaldi and Chromium. Firefox support is planned.
Explore Govern ↗The check runs on the device, with the detectors and the actions your policy sets. The legal AI tool needs no change.
31 built-in detectors for personal identifiers, secrets and API keys, and UK, EU and US identifiers, plus patterns you write.
Monitor, Warning, Redact or Block, set per detector. Redact removes the sensitive span and sends the rest.
Files uploaded to the AI website, read for instructions aimed at the model.
Off, or on at Monitor, Warning or Block. Flagged documents are listed by verdict on the tool's profile.
The AI websites reached from the firm's browsers, Harvey and Legora among them.
Each one is sanctioned, unsanctioned or blocked. The decision is yours, and it is recorded.
Enrol the device. See the use. Set the policy. Then keep the record.
Install the Layer 1 Agent. The browser extension is installed by policy, so nobody has to add it by hand.
Agent installation ↗Visits to Harvey, Legora and the other AI websites the firm's browsers reach, by person and by device.
Shadow AI detection ↗Choose an action for each detector, add your own patterns, and switch on the uploaded document rule.
Govern ↗Each detection is an event with its time, detector, action and confidence, and Events can be filtered to the site it happened on.
Monitor ↗What is reported is the detector, the action, a count, the confidence and the site's domain.
Where your organisation allows evidence, an administrator sees the prompt with every match replaced by the detector's name. A flagged document carries its file name, never its contents.
Each time an administrator opens that evidence, the audit log records it.
Coverage, safety, access, browsers, privilege, confidentiality and the record. These are the details to settle in your evaluation.
Harvey and Legora are built in, each with its own profile page that counts its regional tenants as one product. The same prompt checks apply on the other AI websites the browser extension guards. Ask for coverage of your exact tools.
Harvey AI risks depend on how your lawyers use it, and EndLayer does not certify any vendor as safe. EndLayer shows who uses Harvey, checks what goes into each prompt and flags prompt injection in uploaded documents, so your firm can show its Harvey AI governance: Harvey AI security, Harvey AI data privacy and use by person.
The same controls apply to Legora: Legora security, Legora data privacy and Legora governance in one profile, with use by person, prompt checks and a sanction decision.
No. EndLayer works on the device and in the browser, so it needs no access to the legal AI tool and the tool needs no change. EndLayer is not affiliated with Harvey or Legora, and the names belong to their owners.
Chromium browsers: Chrome, Edge, Brave, Vivaldi and Chromium. Firefox support is planned. The extension is installed by policy.
Not as a category. The built-in detectors look for personal identifiers, secrets and similar patterns. You can add patterns of your own, such as the format of a matter number, and choose what happens when one matches.
Matched text stays on the device. The console records the detector, the action and the count, never the client data itself.
The tool's profile: 30 days of use by person, and how many prompts were blocked, warned on, redacted and logged. Behind it, each detection is an event with its time, detector, action and confidence.
See the Guardian Agent Platform on your own fleet. Find every AI tool, understand what it reaches and costs, and apply the controls your team needs.