Book a demo

Data loss prevention for AI

Security at the source.

Chassis is DLP that runs on your users' devices, right beside each AI app, and stops requests that would leak your data before they leave.

We'll show you Chassis stopping a real request, live.

Exploded view of a user's device running Chassis The user's device is the base. The Chassis frame sits on it, with the Chassis agent at the front. A chat app and an agent container are mounted above the frame on struts. Requests from each rise toward the model APIs; Chassis checks each one at a ring on its strut and stops the one carrying an AWS key.
    Exploded view. Requests and rules shown are examples.Scroll to assemble

    Bolted to the source, not routed through a cloud.

    Most DLP tools for AI send every prompt through the vendor's servers to inspect it. That makes the vendor one more place your most sensitive data ends up. Chassis checks each request on the user's device, in the same place it was written, beside the app that wrote it.

    A cloud DLP proxyEvery prompt takes a detour. The inspection itself creates a new archive to secure, audit and answer subpoenas for.
    ChassisThe check happens before the request leaves. There is no detour and no copy, because there is nowhere for one to go.
    In 2023, Samsung restricted generative AI tools at work after employees pasted internal source code into ChatGPT.Widely reported, including by Bloomberg, May 2023

    Bans don't hold. They push people to personal accounts, where you see even less. Chassis lets them keep using AI and keeps the sensitive parts at home.

    Three answers at the moment of sending.

    People see a clear note right where they typed, never a silent failure or a ticket. In DLP terms: block, redact, or warn and confirm. Your rules decide which answer each request gets.

    ~/acme-api · claude
    > deploy the staging build, use key AKIAIOSFODNN7EXAMPLE
    Stopped before sending

    This prompt contains an AWS access key. Keys stay on your computer. Remove it and try again.

    rule secrets.aws-access-key · policy rev 14

    For things that should never leave.

    Access keys, card numbers, credential files. Chassis stops the prompt, the file read, the shell command or the web fetch, and says exactly why.

    Where
    Claude Code in the terminal
    Caught
    AWS access key in a prompt
    Sent
    Nothing

    Example request

    Rules go down. Nothing comes back.

    You publish a policy and sign it. Each user's device pulls it, checks the signature, and rejects anything unsigned or older than what it already has. That pull is the only request Chassis makes.

    • Prompts and repliesstay with the user
    • Files and pasted textstay with the user
    • Activity logstays with the user
    • Device and user IDsnever collected
    Signed policy flows down from the console to each user's device; nothing flows back up. ADMIN CONSOLE policy.json · rev 14 SIGNED · ED25519 USER USER USER NO UPLOAD PATH

    What we never collect, we can never leak.

    So what can I see?

    The policy you published and which version is live. Install and health status come from your MDM, which already reports what's on each device. Each person can see their own activity on their own machine.

    Specifications

    What Chassis covers today, and what's next. We'd rather you hear the gaps from us.

    SystemTodayNext
    CategoryEndpoint DLP for AI: every check and every decision happens on the user's device—
    ActionsBlock, redact, or warn and let the person confirm—
    AI chatsChatGPT, Claude, Gemini, Copilot, Perplexity, DeepSeek and Mistral in ChromeChatGPT and Claude desktop apps
    Coding agentsClaude Code: prompts, file reads, shell commands and web fetchesCursor and other coding agents
    FilesCredential files and sensitive file contents read by Claude CodeUploads and drag-and-drop in chat apps
    BrowsersChromeEdge and Firefox
    Baseline rulesAPI keys and tokens, card numbers, Social Security numbers, credential files, contact lists, confidentiality markings—
    Your rulesClient names, project codenames, internal hostnames, any term you addSeparate policies per group
    DeploymentAgent and browser extension pushed through Jamf, Intune or Kandji; no proxy, no network rerouting, no certificates—
    Policy deliverySigned bundles, pulled by each user's device; unsigned or older versions rejected—

    Live in an afternoon.

    1. Choose rules

      Start from the baseline, then add the names and codenames that matter to your company.

    2. Deploy

      Ship the agent and the extension with the MDM you already run. Your users' devices pull the signed policy.

    3. Protect

      From the next message on, Chassis checks every AI request your users make and stops, cleans or asks.

    See it stop a real request.

    Tell us about your team. We'll walk you through Chassis live and answer your DLP and security questions.