CirvixArchitecture

System architecture & execution flow.

Every agent action crosses a single control layer before it reaches a model, service, or database. This is what that layer is made of and what happens inside it. It is the enforcement core of runtime security for AI agents.

1control layer, no bypass path
5decisions per execution
1.2mstarget policy evaluation latency
UDSlocal socket, no network hop
Controlled by design

The execution path.

Traffic enters from any runtime, resolves through the core, and only reaches an output once a verdict allows it. This is the execution path our AI agent security model enforces.

01 / Inputs
Client SDKsTyped execution events
Agent FrameworksBring your own runtime
WebhooksSigned external triggers
02 / Core engine
CONTROL ACTIVE

Cirvix Control Plane

Policy Enforcement Engine
Tamper-Evident Audit Logger
AES-256 Envelope Encryption
03 / Outputs
LLM ProvidersScoped model routing
External APIsSanctioned tool calls
DatabasesTenant-aware access
The pipeline

Five decisions,
in this order.

Each stage can refuse. A stage that refuses does not fall through to the next one.

01

Identity — who is asking

Resolve the agent and the org from the credential, not from a field in the request body. An agent cannot claim a different org_id than the one it authenticated as.

02

Scope — what it is allowed to reach

Tenant scoping is applied in routing. A query that is not scoped does not become scoped later; it is refused.

03

Policy — what the rules say

The rule set resolves to exactly one verdict. Explicit deny is terminal, no match means deny, and audit_only never authorizes.

04

Evidence — what gets recorded

The decision is written into the hash-linked chain before the action is released. The record does not depend on the action succeeding.

05

Release — what actually runs

Only an allowed action proceeds, carrying the scope it was granted and nothing wider.

primitive-flow.mdDELEGATION TO PROOF

Delegation → Effective Action → Consequence → Proof

Security capabilities cannot remain separate dashboards that humans visit after a breach. CIRVIX operates as an agent-callable enforcement primitive embedded directly in the execution flow: Agent → CIRVIX Decision (<1.2ms design target) → Tool Invocation → Real-World Side Effect → Hash-Chained Evidence Record (optionally Ed25519-signed). Enforcement is protocol-agnostic, spanning MCP, REST APIs, CLI commands, browser automation, sub-agent swarms, and local OS calls.

Kernel Trapping & Proxies

OS-Level Trapping vs. Proxy Enforcement.

Enterprise security diligence requires clarity on where controls actually bind. Cirvix operates a two-layer defense-in-depth model: high-level semantic proxy inspection plus OS-native kernel enforcement across Linux, macOS, and Windows.

LayerLinuxmacOSWindowsEnforcement Capabilities
Layer 1: Protocol Proxy
stdio / MCP / UDS
Unix Domain Socket (/var/run/cirvix/ipc.sock)Unix Domain Socket (/var/run/cirvix/ipc.sock)Named Pipes (\\.\pipe\cirvix-ipc)Inspects tool payloads, validates schemas, constrains injected agents' tool calls, runs secret DLP scanning, and enforces authorization before execution.
Layer 2: OS Kernel Traps
System Call Containment
eBPF + seccomp-bpf
Hooks sys_enter_execve, sys_enter_connect, and LSM hooks
EndpointSecurity (ES)
Kernel client daemon (ES_EVENT_TYPE_AUTH_EXEC, AUTH_OPEN)
WFP + ETW / Minifilter
Windows Filtering Platform network callouts + ETW process tracing
Contains out-of-band calls outside policy. Even if an agent process is compromised to run raw shell forks or bypass userspace shims, kernel traps deny unauthorized sockets, files, and child processes.
Linux: eBPF & LSM

Kernel Syscall Interception

On Linux hosts and containers, the Cirvix daemon loads compiled eBPF bytecode attached to tracepoints and Linux Security Module (LSM) hooks. Any unbrokered attempt by an agent or spawned child to initiate raw TCP connections or call execve outside declared policy boundaries is dropped directly in the kernel ring before execution.

macOS: Apple EndpointSecurity

Synchronous Auth Events

On macOS, Cirvix utilizes Apple's EndpointSecurity framework (replacing legacy dtrace). The daemon subscribes to synchronous AUTH_EXEC and AUTH_OPEN events, holding process execution until the Cirvix policy engine computes a verdict. Binaries outside policy are killed immediately.

Windows: WFP & ETW

Filtering Platform & Driver Callouts

On Windows workstations and servers, Cirvix binds to the Windows Filtering Platform (WFP) to enforce packet- and connection-level network barriers on agent processes, while Event Tracing for Windows (ETW) and filesystem minifilter drivers provide deterministic audit trails and file tampering locks.

Data path

What crosses
the boundary.

The control plane needs the shape of a request to decide about it. It does not need to keep the contents.

request → decisionPATH
evaluated
agent_id, org_id, action, resource, policy set
recorded
decision, rule id, reason, block hash, prev hash
discarded
payload contents not approved for retention
never stored
CIRVIX_MASTER_KEY — customer managed, no escrow path

Encryption is AES-256-GCM envelope encryption. The master key stays on your side of the boundary; Cirvix holds no escrowed copy and therefore cannot decrypt on your behalf.

Failure behaviour

What happens
when it breaks.

A control plane that fails open is not a control plane. These are the states it is designed to hold.

ConditionBehaviourRecorded
Policy engine unreachableActions are denied. There is no pass-through mode.Denial recorded with cause
Audit store unavailableActions are denied — a decision that cannot be recorded is not released.Recorded on recovery
Rule set fails to parseThe previous valid rule set stays active; the broken one is rejected at load.Load failure recorded
Unknown actionDenied as unmatched. Absence of a rule is not permission.Unmatched decision recorded

Run it inside your own boundary

Docker Compose, Helm, or bare Node — wherever the trust boundary already is.

Deployment models
Govern what ships

Bring every agent
under control.

Set durable policy, preserve a verifiable record, and give teams a safer way to put intelligent systems to work.

Copied to clipboard