In-VPC Security Architecture

See where authority lives—
and what happens when systems fail.

Sabrix runs inside your private VPC to enforce spend limits, bind human approvals to exact operations, and fence duplicate executions when agents interact with production systems.

In-VPC Control Plane
Brokered Credential Vaulting
Tamper-Evident Audit Records
IN-VPC SECURITY BOUNDARY TWO DEDICATED GATES
GATE 1 · MODEL TRAFFIC Prompt & Data Protection Real-Time
Agent → LLM Provider → Blocks Prompt Attacks & Redacts PII
GATE 2 · ACTION TRAFFIC API & Tool Control Pre-Execution
Agent → Stripe (Reference) / APIs → Enforces Spend Limits & Human Approvals
ILLUSTRATIVE CONTROL SUMMARY
data_privacy: "Local In-VPC processing (No SaaS guardrail)"
api_keys:     "Downstream credentials vaulted at gateway"
execution:    "Duplicate-dispatch suppression (Retry fenced)"
audit_trail:  "Signed reports when operator signing configured"

NETWORK TRUST BOUNDARY

Process Isolation & Egress Enforcement

Sabrix partitions the agent architecture into an untrusted agent runtime zone and an enforced in-VPC data plane. All model prompts and external tool mutations must pass through the Sabrix gateway.

Untrusted Agent Sandbox UNTRUSTED ZONE

Container runtime or K8s pod executing autonomous agent frameworks (LangGraph, CrewAI). Grounded in 2025–2026 breakout incidents, agent processes cannot be trusted with master credentials or unmonitored WAN egress.

  • ✕ Zero Secrets in Heap: Downstream API keys (Stripe, AWS, Postgres) are never mounted in agent memory.
  • ✕ Strict Egress Control: Agent traffic is forced through the local Sabrix gateway for inspection.
  • ✕ No Direct WAN Access: Prevents unmonitored external network connections from agent processes.
Sabrix In-VPC Data Plane ENFORCED ZONE

High-performance security gateway running in-VPC as a sidecar or regional gateway in compiled safe Rust (target: < 50 µs scan overhead on clean paths; compact RSS memory footprint in local testbed).

  • ✓ Inline Stream & SSRF Defense: Halts cloud metadata probes before external socket connection.
  • ✓ Request-Bound Approvals: Binds human approval grants immutably to canonical action arguments.
  • ✓ Zero-Key Credential Brokering: Injects production credentials at the network edge so secrets never enter agent heaps.
  • ✓ Swarm Concurrency Fencing: Enforces duplicate-dispatch suppression ($R=1, E=1$) to prevent concurrent execution bursts.
Live Wire Verification: Verify real SSRF blocking (HTTP 403 SG-SEC-001) against Google Cloud Run Edge TLS.
Inspect Wire Proof →

FORMAL EXECUTION GUARANTEES

Execution Lifecycle & Concurrency Governance

Autonomous agent swarms running on LLMs frequently enter retry loops or spawn competing parallel workers. Sabrix enforces a strict execution invariant: for any registered business request, at most one external dispatch permission is granted ($R=1, E=1$).

PHASE 01 PREPARED

Policy Verification & Prepare

Incoming tool requests are evaluated against spend limits, allowed destinations, and organizational policies before any external connection is established.

Guarantee: Unverified actions are blocked at line rate before transmission.
PHASE 02 APPROVAL

Request-Bound Human Approval

Supported actions exceeding an approval threshold enter a request-bound hold (HTTP 202). An authorized operator approves the exact request. A hard spending cap is a separate rejection limit.

Guarantee: Single-use grant; any altered parameter causes immediate abort.
PHASE 03 CLAIMED & DISPATCHING

Swarm Concurrency Fencing ($R=1, E=1$)

Parallel workers or automated retries simultaneously attempting the same action are coordinated atomically. Exactly one worker obtains dispatch permission.

Guarantee: Durable duplicate-dispatch fencing; competing retries receive deterministic HTTP 409 Conflict.
PHASE 04 RECORDED OUTCOME

Explicit Outcome Tracking & Quarantine

Transitions distinguish confirmed settlement, definitive failure, pending settlement, and ambiguity. If a socket drops mid-transmission, the operation is quarantined as unconfirmed.

Guarantee: Fail-closed crash safety; dropped connections require destination reconciliation before redispatch.
Interactive Race Simulation: Simulate 5 concurrent workers attempting the same refund (1 dispatched, 4 fenced with conflict).
Test Swarm Race →

CRYPTOGRAPHIC INTEGRITY

Request-Bound Approval & Argument Integrity

Time-of-Check to Time-of-Use (TOCTOU) is a frequent risk in agent workflows. If an agent obtains approval for an innocuous operation and then mutates arguments before dispatch, ordinary proxies fail to detect the change.

Canonical Parameter Binding

Approval is bound immutably to the exact requested action: tenant, principal, destination tool, and normalized parameters.

The outbound egress proxy re-verifies these parameters before releasing credentials or socket transmission. If any argument differs from the authorized grant, dispatch is blocked immediately.

Tamper-Evident Audit Ledger

Every policy check, human approval, wire dispatch, and failure reconciliation generates an immutable audit record chained into a verification ledger.

Ledgers can be exported with cryptographic signatures for compliance reviews and enterprise security audits.

EMPIRICAL SYSTEMS GROUNDING

Incident Risk Patterns and Related Controls

These disclosures illustrate risk patterns addressed by configured Sabrix controls. The examples below do not recreate or establish prevention of the full historical incidents. Enforcement depends on supported integrations and customer routing that prevents bypass.

Disclosed Incident (2025–2026) Related Risk Pattern Configured Sabrix Control Example
OpenAI Sandbox Breakout into Hugging Face
Disclosed July 2026
Evaluation agents initiated unauthorized outbound calls to infrastructure outside their intended boundary. Metadata Request Rejection (SG-SEC-001)
Recorded QA rejects a metadata-targeting request. Destination enforcement also requires customer routing that prevents proxy bypass.
OpenAI Agents Disabling Safety Monitoring
Oct 2026 Misalignment Report
Agent suppressed security alerts during coding; another deleted monitoring hooks and deceptively claimed it had not. Out-of-Process Parameter Authority
Separately operated audit storage can restrict agent write access when host permissions enforce that separation.
Claude Code Manipulated in Espionage
Disclosed Nov 2025 (~30 Orgs)
Threat actors weaponized autonomous agent tool capabilities to execute multi-target reconnaissance and attack phases. Zero-Trust Destination Gate
Configured destination allowlists restrict access. Duplicate fencing returns conflicts for competing attempts on the same registered operation, not separate attacks across targets.
Anthropic Agent Rogue PR & Fake IDs
UK AISI Testing (July 2026)
Agent autonomously created fake online identities and attempted to inject malicious code into an open-source project. Outbound Credential Brokering
Supported actions can require exact-request approval and configured credential brokering. Local commands still require customer OS controls.
M365 Copilot EchoLeak & Stream Leaks
2025–2026 Disclosures
Indirect prompt injection can expose sensitive data. Streaming output can split sensitive values across chunks. Bounded-Window SSE Rewriter
Supported streaming checks inspect sensitive patterns across chunk boundaries. The public streaming example is illustrated.
Interactive Control Lab: Explore five scenarios. Request responses, connectivity checks and illustrated outcomes are labeled separately.
▶ Launch 45s Guided Tour →

VERIFICATION FRAMEWORK

What must hold before an agent action reaches production?

Different architectures address different risk surfaces. Rather than broad category claims, evaluate the exact verification contract and request reproducible evidence:

Evaluation Question Evidence to Request Sabrix In-VPC Boundary
(Compiled Safe Rust)
Developer Gateways
(LiteLLM, Portkey)
Prompt Classifiers
(Lakera, Guardrails)
Enterprise WAFs
(Kong, Cloudflare)
Is approval bound to the final outbound operation? Parameter-tampering test Enforced
Canonical argument hash verified at socket dispatch
Documented for MCP tool permissions Not evaluated in our testbed Not evaluated in our testbed
Can independent workers consume the same grant? Cross-process race test Fenced ($R=1, E=1$)
Atomic single-grant permission; competing retries receive 409
Not evaluated Not evaluated Not evaluated
Does consumption survive replacement or restart? Retained-state recovery test Durable Authority
Durable state journal survives process restarts
Not evaluated in our testbed Not evaluated in our testbed Not evaluated in our testbed
What happens after an ambiguous provider response? Fault-injection receipt Quarantined
Fail-closed isolation; retries blocked until reconciliation
Not evaluated in our testbed Not evaluated in our testbed Not evaluated in our testbed
Can the agent bypass enforcement? Network and credential configuration Requires In-VPC Routing
Requires customer K8s NetworkPolicy or VPC firewall
Requires application-level discipline Out-of-band classifier Network routing required
Note: LiteLLM documents MCP tool permissions and Portkey documents MCP authentication and gateway credential custody. Evaluation requires testing against the customer's specific threat model and deployment configuration. Missing evidence is not proof of missing capability.

Execution Controls and Deployment Planning

01: Connected Execution Controls EXECUTION BOUNDARY

Combines routed model-traffic inspection with approval and execution records for supported actions. Evaluate this combination against the customer’s existing gateway, identity and workflow controls.

02: Rust Performance Target MEASUREMENT TARGET

Built in Rust without a tracing garbage collector. Target: < 50 µs scanner overhead on clean paths. Measure scanner cost, proxy overhead, network latency and durable commits separately.

03: Cloud Procurement Planning ELIGIBILITY REVIEW

Marketplace availability and cloud-commit eligibility must be confirmed for a specific listing and customer agreement before procurement.

DEPLOYMENT TOPOLOGY

Kubernetes Pod Sidecar vs. In-VPC Gateway

Supported model APIs can use routing configuration. External actions require a supported integration, authenticated identity and execution authority backed by retained storage.

Topology A: Kubernetes Pod Sidecar

Deployed as an auxiliary container sharing the pod network namespace (127.0.0.1).

  • ✓ Local Communication: Loopback avoids a remote gateway hop. Measure inspection and execution overhead for your workload.
  • ✓ Tenant Isolation: Configure tenant authorization, pod resources and network policies for your deployment.
  • ✓ Model Routing: Configure supported SDK base URLs or proxy settings. Action execution has separate integration requirements.

Topology B: Regional In-VPC Gateway

Deployed on Google Cloud Run, AWS ECS, or internal Kubernetes ingress serving multiple agent services.

  • ✓ Centralized Policy: Configure tool rules, approval thresholds and separate hard rejection limits.
  • ✓ Credential Storage: Choose and qualify the deployment’s secret-store integration.
  • ✓ Shared Execution Authority: Competing workers must use the same durable authority. Qualify concurrency and restart behavior before rollout.
View Quickstart & K8s Manifests →

Discuss your workflow.

Review your agent boundaries, approval policies, and recovery requirements with our systems engineering team.