Raziel: Agent Authorisation at Superhuman Scale

We built Raziel to give AI agents reusable, proof-backed permissions bound to an exact agent and task, with independent verification and private relationship evidence kept by the data owner.

poster

Description

Raziel: Agent Authorisation at Superhuman Scale

A task is not a grant. An AI agent can be asked to review a project. The service holding Doc 42 still needs to know whether that agent is an assigned reviewer.

Why this matters

AI agents need to work across teams and services. Each service has its own rules for what an agent can read. A signed access token lets the receiving service check who issued a grant. The service still relies on the issuer to have applied the access rule correctly. We built Raziel so the receiving service can independently check that the grant follows the approved rule.

What we built and how it works

Raziel gives an agent a reusable grant tied to its identity, task, and data scope. The grant carries a zero-knowledge proof that the access request met an approved rule.

  • Compile the rule. Raziel converts supported SpiceDB rules into a policy that a Rust program running in SP1 can evaluate.
  • Prove the decision. The program checks the policy against a committed snapshot of relationships, such as team memberships. The prover generates a proof for an allowed request. The private relationship records stay with the data owner.
  • Verify before release. A separate verifier checks the proof without contacting the issuer. The agent can reuse the grant within its scope and validity period. The gateway checks the signature and scope of each request before releasing data.

What we tested

In our demo, one agent reads synthetic payment-processor records. A second agent uses its own grant to check synthetic bank settlements. They exchange a settlement result, and audit receipts trace the protected reads.

  • 447 ms for one independent proof check, with the issuer and prover stopped.
  • About 16 minutes to generate one proof locally in a separate run.
  • 8/8 tampered inputs rejected.
  • 21/21 policy decisions matched SpiceDB on the test cases.

Raziel is the authorisation foundation for Hypha, our broader research into how agents find relevant data and control what they share.

Explore Raziel and work with us

Explore Raziel and discuss a pilot
We’re looking for design partners at NUS FinTech Lab. Bring one action you’d like to delegate to an agent and one access rule it must follow.

Vote for MCOMP-04 at 29th STePS
Visit MCOMP-04 to see the demo. On UVents, select MComp, then Raziel.