Open dashboard

Local proxyZero-knowledge attestationVerified on Solana

Prove an agent's action was screened
without showing what it was.

Blindseal is a privacy-preserving, verifiable screening layer for AI agents.
Nothing sensitive leaves the machine.

npx blindseal init
Read how it works
  • Local-only dashboard
  • Single-use attestations
  • Content-free audit log

01 / The problem

AI agents are easy to trick — put the wrong text in front of one and it does what the attacker wants.

When a tricked agent shows up at an app, the app can't tell.

The only fix today is a service that reads all your agent's prompts, so proving you're safe means handing over your secrets.

02 / How it works

Screen, attest, prove, verify.

Blindseal checks every action your agent takes, on your own machine, and never sends your data anywhere. When your agent needs to prove it was checked, Blindseal gives it a one-time stamp.

  1. Screen

    It sits between an agent and its tools as a local proxy, inspecting every tool call with four layers of prompt-injection detection before the call executes.

  2. Attest

    Blindseal produces a zero-knowledge attestation that a specific tool call passed a committed policy.

  3. Prove

    The proof is generated locally. It is single-use.

  4. Verify

    Verified with one Solana instruction, so any program can gate agent actions without seeing the prompt, the arguments, the operator's rules, or which agent was screened.

Read the long version

03 / Hidden vs public

What stays on the machine,
and what the app sees.

The app sees the stamp and nothing else — not what the agent was asked, what it did, or who it is. Each stamp works once.

Hiddennever leaves the machine

  • Prompt and tool arguments
  • Policy contents, thresholds, custom lists
  • Which agent / operator was screened
  • Model-layer scores

Publicon the attestation

  • That a valid attestation exists
  • Policy root (a hash)
  • Detector version
  • Action hash, expiry, nullifier

What we claim

Blindseal proves that a specific call was screened under a committed policy by a specific detector version, and that the deterministic layers passed. The model-based layers run locally or in a TEE and are attested, not proven. Blindseal does not prove an agent is safe, and does not cover traffic that doesn't pass through the proxy.

Screen the next tool call.

npx blindseal init