In development · proposed architecture

SLM FloodGate™ 2.0

Action-level security for AI agents. FloodGate 2.0 is being designed to help organisations control what an AI agent can do before an action reaches a protected system.

What we are working to achieve

As AI agents move from answering questions to taking actions, organisations need control over the actions those agents can perform across APIs, services and private networks. FloodGate 2.0 aims to place an authorisation boundary between an agent's request and a protected action.

Design objective: a protected action should proceed only when it has the required current, specific authorisation. This is an engineering target subject to implementation, testing and the security of the deployment.

FloodGate 2.0 is a proposed next-generation architecture. It is under development and is not presented as fully built, independently verified or ready for enterprise deployment.

The intended workflow

An AI agent's request moves through a controlled path before it can reach a protected resource. The steps below describe the architecture we are working towards.

AI agent request · API, MCP or private network
↓
Admission controls · manage incoming requests
↓
Risk intelligence · advisory assessment
↓
FloodGate control plane · identity, policy and authorisation
↓
Protected execution · verify and carry out the authorised action
↓
Protected system or resource

1. Request and admission shield

An agent requests an action through an API, MCP connection or private network. The admission shield is intended to limit connections, request rates and processing work before traffic reaches the trusted core.

2. Non-LLM AI risk advisor

The proposed advisor analyses unusual behaviour and potential coordination between agents, including swarm activity. It can recommend increased scrutiny or restriction, but cannot grant authority or override a denial.

3. FloodGate control plane

The proposed deterministic control plane checks identity, session, freshness, policy and the exact requested action. Its intended output is a short-lived authorisation bound to that action. A request that fails the required checks should be denied.

4. Protected execution and resource

The protected execution plane is intended to verify the authorisation, consume it once and carry out only the matching operation. It holds the downstream credentials; the architecture aims to prevent agents reaching the protected resource by a direct bypass.

5. Deterministic watcher

The proposed watcher monitors runtime activity and policy or budget violations, feeding a response back to the control plane. It can maintain or reduce authority through restriction, quarantine or revocation; increasing authority requires a new authenticated authorisation.

Handling agent swarms

A swarm is a group of agents acting or sending requests at the same time. FloodGate 2.0 is being designed to address both heavy request volumes and potentially coordinated behaviour, while continuing to evaluate each protected action against the applicable authority.

Bound incoming work

The admission shield is intended to limit connections, request rates and queues, with backpressure or throttling when capacity is reached. This aims to stop a surge from creating unlimited work for the trusted core.

Detect coordination risks

The non-LLM risk advisor is intended to examine patterns across agents for unusual behaviour and possible coordination. Its findings can trigger scrutiny or restriction, but cannot approve an action.

Keep authority and budgets bounded

Proposed policy and budget controls limit the work and authority available across agents, so creating more agents is not intended to multiply the permissions or resources available to a shared source.

Respond as conditions change

The watcher is intended to feed observed limits and violations back to the control plane, which can restrict, throttle, quarantine or revoke access in line with policy.

Swarm resilience is a development and testing objective. Limits, detection effectiveness and behaviour under load still need to be measured and independently challenged.

Security by design

Specific authority

Design authority around the action and resource required, rather than giving an agent broad ongoing permissions.

Deterministic control

Keep authorisation independent of probabilistic AI. The design target is that if the non-LLM risk advisor is unavailable or wrong, deterministic controls remain active and do not fail open. No LLM is required in the proposed security path.

Protected execution

Work towards a mediated path to protected resources, with deployment controls designed to prevent agents bypassing that path.

Evidence and assurance

Build and challenge the architecture through implementation, adversarial testing, and independent review as development progresses.

These are design objectives, not a guarantee of perfect detection or security. Any future assurance claim will depend on evidence and clearly stated deployment assumptions.

Designed to work with existing security

FloodGate 2.0 is intended to complement identity, endpoint, network and security monitoring systems. These tools can provide identity, telemetry and containment inputs to the control plane. Target environments include online services and private enterprise networks, including API, MCP and internal machine-to-machine paths.

EDR / XDR
SIEM
Identity platforms
Network controls
APIs / MCP

FloodGate 2.0 is a proposed architecture under development. Its planned capabilities and deployment approach remain subject to implementation and validation.

Discuss FloodGate →