Kyvvu

Data leak prevention

Read the record or send the email, but not both in one task.

Agents leak data through ordinary steps: a lookup, a summary, a message. Kyvvu remembers what a task has read, and blocks outbound sends once sensitive data is in play, before they run.

support-agent · task 7f3akyvvu engine · in-process
  1. step.messagePOSTsmtp/externalblock
Blocked: outbound send after personal data was read in this taskTry it: click the highlighted step to take it off the path

The problem

Agent leaks look like ordinary work.

  1. 01

    Every step is allowed on its own

    The agent is allowed to read customer records and allowed to send email. The leak is the combination, in one task.

    path-dependent risk

  2. 02

    Content scanning misses summaries

    An agent rarely pastes the raw record. It summarises, rephrases, translates. Pattern-matching on content sees nothing it recognises.

    DLP blind spot

  3. 03

    Injection makes it targeted

    A hidden instruction in a document can ask the agent to send exactly what it just read to exactly where the attacker wants it.

    indirect prompt injection

In practice

Same send, different verdict, depending on what came before.

Customer record, then an outside email

A support agent looks up a customer to answer a question. A crafted message asks it to forward 'the account details' to another address.

support-agent · task 7f3akyvvu engine · in-process

Illustrative scenarios. Each one is decided by a single policy from the open policy library, as the engine evaluates it.

How Kyvvu helps

Stops the leak without stopping the work.

  1. 01

    Judges the whole path

    Every action is checked against everything the agent did earlier in the task, and blocked before it runs if the sequence is dangerous.

    allow/warn/block

  2. 02

    Driven by your data classification

    A read becomes sensitive because of how the data is classified, e.g. PII. The rule follows the data, whichever tool or channel comes next.

    data.classification: pii

  3. 03

    Credentials too

    Another policy in the same pack blocks any message after a credential has been read in the task.

    credential exfiltration

  4. 04

    Runs inside your environment

    The engine sits in the agent's own process: so there's no proxy or extra network hop, and a check takes under a millisecond. Prompts and data stay with you.

    in-process/no proxy

Questions

Common questions.

01

What counts as sensitive?

Whatever you classify as sensitive when the agent reads it, such as PII. The policy is a few lines of YAML and you can change the classification it reacts to.

02

Does Kyvvu see our data?

No. The engine runs inside the agent's process and decides there. If you'd rather keep traces on your own machines too, one setting does that.

03

What about outbound domains?

The data exfiltration guard pack also restricts outbound calls to an approved domain list, and can require a review step before any external message.

Get started

Every agent has to ASK before it acts.

Start with one agent

See it on your own agents

A 30-minute demo with someone who builds Kyvvu: your agent, your rules, and the policy that decides each step.

Try it now

Watch the verdict flip in your browser

Build a path in the playground, add a PII read, and see the send go from ALLOW to BLOCK.

Want to see other teams and sectors? All use cases.