In design & verificationSecurity gateway for AI agents

Check, decide, and record — before an AI agent reaches your systems.

Graatii is building a security gateway that sits between AI agents and a company's tools, data, and APIs.

At a glance

Position
Between agents and tools, data, APIs
Function
Identity · Policy · Credential brokering · Revocation · Audit evidence
Default
Deny anything not confirmed
Now
Design and verification. No product has shipped.

01Problem

Once agents do real work, access becomes the question

AI agents send email, edit files, and call payment APIs on someone's behalf. The question a company has to answer shifts from how capable the model is to who allowed this request, and how far.

  1. Whose intent is unclear

    An agent acts for a person, but the tool on the other end can rarely tell which user a request serves, or which task it was allowed for.

  2. Secrets live too long and reach too far

    API keys and tokens pasted into agent configuration end up used more broadly, and for longer, than any single task needs.

  3. What an agent reads can change what it does

    Instructions hidden in documents, web pages, or tool descriptions can redirect an agent's next action (prompt injection). The model's own judgment is not a reliable filter for this.

  4. What happened is hard to prove

    Afterwards it is hard to show, with evidence, what was allowed, what was blocked, and whether any path stayed open after access was revoked.

02Approach

Between agents and systems, deciding each request

The Graatii gateway stands where agent requests travel to tools, data, and APIs. What follows is the behavior we are designing and verifying now.

How a request passes through the Graatii gateway A request from an AI agent enters the Graatii gateway, where identity is verified, a policy decision is made, a credential is brokered, and evidence is recorded. Only allowed requests are delivered to the company's tools, data, and APIs; held requests are delivered after a person approves; denied requests stop. Credentials are lent per call from a separate vault, and decisions and deliveries are written to an evidence record. A direct route that skips the gateway is tested to confirm it is closed, and reported as unmanaged if that cannot be confirmed. Requesting side AI agents In-house agents Cloud agents Developer-tool agents Request Graatii gateway 1 Verify identity 2 Decide by policy 3 Broker credentials 4 Record evidence Allow — delivered Hold — waits for approval Deny — stopped Your systems Tools, data, APIs MCP servers Internal APIs SaaS Credential vault Separate zone Lent per call Evidence record Decisions, delivery and the basis A direct route that skips the gateway Tested to confirm it is closed; otherwise reported as unmanaged How a request passes through the Graatii gateway A request from an AI agent enters the Graatii gateway, where identity is verified, a policy decision is made, a credential is brokered, and evidence is recorded. The decision is allow, hold, or deny, and only allowed requests reach the company's tools, data, and APIs. Credentials are lent per call from a separate vault, and decisions and deliveries are written to an evidence record. A direct route that skips the gateway is tested to confirm it is closed, and reported as unmanaged if that cannot be confirmed. Requesting side AI agents In-house agents Cloud agents Developer-tool agents Request Graatii gateway 1 Verify identity 2 Decide by policy 3 Broker credentials 4 Record evidence Allow Hold Deny Credential vault Separate zone Lent per call Evidence record Decisions, delivery and the basis Allowed requests only Your systems Tools, data, APIs MCP servers Internal APIs SaaS A route that skips the gateway Tested to confirm it is closed; otherwise reported as unmanaged.
Figure 1. A request reaches its destination only after it is checked and decided at the gateway. Routes that skip the gateway are tested separately to confirm they are closed.
  • Identity

    Confirms which agent is asking, on whose behalf, and for which task, and carries that identity with the request. High-risk requests whose link to a user and task cannot be verified are not sent.

  • Policy decision

    Decides each request before it runs: allow, deny, or hold for approval. The final basis is a policy people can read; model output is an input, never the verdict.

  • Credential brokering

    Agents are not handed long-lived secrets. The gateway lends a short-lived credential for one approved call to one destination, and keeps secrets in a separate zone.

  • Revocation

    When access is revoked, it confirms that open connections and issued credentials are cut off, not only new requests. Until then, the status reads "revocation in progress".

  • Audit evidence

    Records what was decided, whether it was actually delivered, and on what basis. Protection is reported only as far as that evidence supports.

03Principles

Four principles that hold the design

Every addition to the design is measured against these four.

  1. Deny by default

    Only confirmed permission passes. Ambiguous or unparseable requests are rejected, unregistered operations never fall through to allow, and high-risk requests stay closed when evaluation fails.

  2. "Done" only with evidence

    Revocation is marked complete only after evidence shows every path is closed. A blocked new connection is not enough, because an already-open connection can still carry the old permission.

    Protection labels work the same way: without evidence, we do not say "secure".

    1. Revocation requested

      A request to withdraw access arrives

    2. Revocation in progress

      Each path is checked

      Paths to confirm

      • New connections
      • Already-open connections
      • Issued credentials
      • Restarted servers
    3. Revocation complete

      Only after evidence shows every path is closed

    Figure 2. When the revocation status may change. If any path is unconfirmed, it stays "in progress".
  3. Independent verification

    We check the design repeatedly, with different reviewers and with measurements. We also run tests against deliberately broken setups to see that they really catch failures. When we find a case that looks like a pass but still leaks access, it becomes a rule and a test.

  4. Complexity below the surface

    The depth of decisions and evidence stays inside the system. Where people look, the risk, the reason, the impact, and the next step should be readable at a glance.

04Status

We are in design and verification

Each stage is judged by its own evidence, and evidence from one stage is never used to claim the next. No product has shipped yet.

  1. 1Now

    Design & verification

    Security contracts and trust boundaries are written down and checked through independent review and small measurements.

  2. 2

    Build

    Built to those contracts and checked against acceptance tests defined in advance.

  3. 3

    Independent verification

    Someone other than the builder reruns the same tests under the same conditions and gets the same result.

  4. 4

    Production readiness

    Only after operational criteria such as failure handling and tenant isolation are confirmed does it go into a customer environment.

05Contact

Reach us by email or phone