SINT Protocol

Formally specified security, permission, and economic enforcement layer for physical AI.

SINT is the missing governance layer between AI agents and the physical world. Every tool call, robot command, and actuator movement flows through a single Policy Gateway that enforces capability-based permissions, graduated approval tiers, and tamper-evident audit logging.

Academic grounding: SINT is formally specified against IEC 62443 FR1–FR7, EU AI Act Article 13, and NIST AI RMF. The formal specification and evaluation framework reference the ROSClaw empirical safety study (arXiv:2603.26997) and MCP security analysis (arXiv:2601.17549).

Agent ──► SINT Bridge ──► Policy Gateway ──► Allow / Deny / Escalate
                               │
                       Evidence Ledger (SHA-256 hash-chained)
                               │
                    TEE ProofReceipt (Intel SGX / ARM TrustZone)

Why SINT?

AI agents can now control robots, execute code, move money, and operate machinery. But there's no standard security layer between "the LLM decided to do X" and "X happened in the physical world."

The empirical case for SINT:

ROSClaw (IROS 2026): Up to 4.8× spread in out-of-policy LLM action proposals across frontier models under identical safety envelopes. The 3.4× divergence between frontier backends is measurable, reproducible, and persistent.
MCP security (arXiv:2601.17549): 10 documented real-world MCP breaches in under 8 months, including a CVSS 9.6 command injection affecting 437,000 downloads.
SROS2: Formally demonstrated to contain 4 critical vulnerabilities at ACM CCS 2022, including access-control bypasses permitting arbitrary command injection.
Unitree BLE worm (September 2025): Hardcoded crypto keys enabled wormable BLE/Wi-Fi command injection across robot fleets — precisely the scenario SINT's per-agent token scoping and real-time revocation prevent.

Core guarantees:

No agent action ever bypasses the Policy Gateway (invariant I-G1: No Bypass)
Every decision is recorded in a tamper-evident SHA-256 hash-chained ledger (invariant I-G3: Ledger Primacy)
Physical constraints (velocity, force, geofence) are enforced at the protocol level — in the token, not in config
E-stop is universal across all non-terminal DFA states (invariant I-G2: E-stop Universality)
Per-agent capability tokens with real-time revocation

Quick Start

# Prerequisites: Node.js >= 22, pnpm >= 9
pnpm install
pnpm run build
pnpm run test        # 815 passing tests across 30 workspace members

Start the Gateway Server

pnpm --filter @sint/gateway-server dev
# → http://localhost:3100/v1/health

Run a Single Package

pnpm --filter @sint/gate-policy-gateway test
pnpm --filter @sint/bridge-mcp test

For AI Agents

If you are an AI agent (Claude, GPT, Gemini, Cursor, etc.) working in this repo, read AGENTS.md first. It covers key invariants, common mistakes, and entry points for the most common tasks. For deeper implementation details, see CLAUDE.md.

Architecture

┌──────────────────────────────────────────────────────────────┐
│  AI Agents / Foundation Models                               │
│  (Claude, GPT, Gemini, open-source)                         │
└──────────────────┬───────────────────────────────────────────┘
                   │
┌──────────────────▼───────────────────────────────────────────┐
│  SINT Bridge Layer (L1)                                      │
│  ┌────────────┐ ┌────────────┐ ┌────────────┐ ┌──────────┐  │
│  │ bridge-mcp │ │ bridge-ros2│ │ bridge-a2a │ │ bridge-  │  │
│  │ MCP tools  │ │ ROS topics │ │ Google A2A │ │ open-rmf │  │
│  └────────────┘ └────────────┘ └────────────┘ └──────────┘  │
│  ┌──────────────────────┐ ┌───────────────────────────────┐  │
│  │ bridge-mqtt-sparkplug│ │ bridge-opcua                  │  │
│  │ Industrial IoT       │ │ PLC / OT control plane bridge │  │
│  └──────────────────────┘ └───────────────────────────────┘  │
│  Per-resource state: UNREGISTERED→PENDING_AUTH→AUTHORIZED    │
│  →ACTIVE→SUSPENDED (real-time revocation without restart)    │
└──────────────────┬───────────────────────────────────────────┘
                   │ SintRequest (UUIDv7, Ed25519, resource, action, physicalContext)
┌──────────────────▼───────────────────────────────────────────┐
│  SINT Gate (L2) — THE choke point                           │
│  ┌─────────────────────────────────────────────────────────┐ │
│  │  PolicyGateway.intercept()                              │ │
│  │  1. Schema validation (Zod)                             │ │
│  │  2. Token validation (Ed25519 + expiry + revocation)    │ │
│  │  3. Resource scope check                                │ │
│  │  4. Per-token rate limiting (sliding window)            │ │
│  │  5. Physical constraint enforcement                     │ │
│  │  6. Forbidden action sequence detection                 │ │
│  │  7. Tier assignment: max(BaseTier, Δ_human, Δ_trust...) │ │
│  │  8. T2/T3 → escalate to approval queue                 │ │
│  │  9. T0/T1 + approved T2/T3 → allow                     │ │
│  │  10. Bill via EconomyPlugin (if configured)             │ │
│  └─────────────────────────────────────────────────────────┘ │
│                          ↓                                   │
│  EvidenceLedger (SHA-256 hash chain + TEE ProofReceipt)     │
└──────────────────────────────────────────────────────────────┘

APS vs SINT Primitives

APS Concept	SINT Implementation
Principal	`agentId` (Ed25519 public key) + W3C DID identity
Capability	`SintCapabilityToken` (Ed25519-signed, scoped, attenuatable)
Authority	`PolicyGateway.intercept()` — single choke point
Confinement	Per-token resource scope + physical constraints (velocity, force, geofence)
Revocation	`RevocationStore` + ConsentPass endpoint (real-time)
Audit	`EvidenceLedger` — append-only, SHA-256 hash-chained

Packages

Gate (Security Core)

Package	Description	Tests
`@sint/core`	Types, Zod schemas, tier constants, formal DFA states	—
`@sint/gate-capability-tokens`	Ed25519 tokens, delegation, W3C DID identity	55
`@sint/gate-policy-gateway`	Authorization engine: tiers, constraints, rate limiting, M-of-N quorum	152
`@sint/gate-evidence-ledger`	SHA-256 hash-chained append-only audit log with TEE attestation	45

Bridges (11 bridges)

Package	Description	Tests
`@sint/bridge-mcp`	MCP tool call interception and risk classification	66
`@sint/bridge-ros2`	ROS 2 topic/service/action interception with physics extraction	20
`@sint/bridge-a2a`	Google A2A Protocol bridge for multi-agent coordination	38
`@sint/bridge-iot`	Generic MQTT/CoAP edge IoT bridge with gateway session interception	21
`@sint/bridge-mqtt-sparkplug`	MQTT Sparkplug profile mapping with industrial command tiering defaults	8
`@sint/bridge-opcua`	OPC UA node/method mapping with safety-critical write/call promotion	6
`@sint/bridge-open-rmf`	Open-RMF fleet/facility mapping for warehouse dispatch workflows	5
`@sint/bridge-economy`	Economy bridge: balance, budget, trust, billing ports	47
`@sint/bridge-mavlink`	MAVLink drone/UAV command bridge	15
`@sint/bridge-swarm`	Multi-robot swarm coordination bridge	9
`@sint/bridge-economy`	Economy enforcement: balance, budget, trust, billing	47

Engine (AI Execution Layer)

Package	Description	Tests
`@sint/engine-system1`	Neural perception: sensor fusion, ONNX inference, anomaly detection	42
`@sint/engine-system2`	Symbolic reasoning: behavior trees, task planning, System 1/2 arbitration	86
`@sint/engine-hal`	Hardware Abstraction Layer: auto-detect hardware, select deployment profile	26
`@sint/engine-capsule-sandbox`	WASM/TS capsule loading, validation, and sandboxed execution	36
`@sint/avatar`	Avatar Layer (L5): behavioral identity profiles, CSML-driven tier escalation	25

Reference Capsules

Package	Description	Tests
`@sint/capsule-navigation`	Waypoint following navigation reference capsule	11
`@sint/capsule-inspection`	Visual anomaly detection for manufacturing QA	8
`@sint/capsule-pick-and-place`	Gripper control for pick-and-place tasks	12

Persistence

Package	Description	Tests
`@sint/persistence`	Storage interfaces + in-memory/PG/Redis implementations	26
`@sint/persistence-postgres`	Production PostgreSQL adapters for ledger, revocation, and rate-limit durability	14

Apps & SDKs

Package	Description	Tests
`@sint/gateway-server`	Hono HTTP API with approvals, SSE streaming, A2A routes	—
`@sint/mcp`	Security-first multi-MCP proxy server	—
`@sint/dashboard`	Real-time approval dashboard with operator auth	29
`@sint/client`	TypeScript SDK for the Gateway API (delegation, SSE)	—
`@sint/sdk`	Zero-dependency public TypeScript SDK aligned to gateway v0.2 contracts	9
`@sint/conformance-tests`	Security regression suite — all phases	—

Total: 30 workspace members, 815 passing tests

Approval Tiers

Graduated authorization mapped to physical consequence severity:

Tier	Name	DFA States	Auto-approved?	Example
T0	OBSERVE	→ OBSERVING	Yes (logged)	Read sensor data, query database
T1	PREPARE	→ PREPARING	Yes (audited)	Write file, save waypoint, stage plan
T2	ACT	ESCALATING → ACTING	Requires review	Move robot, operate gripper, publish `/cmd_vel`
T3	COMMIT	ESCALATING → COMMITTING	Requires human + optional M-of-N	Execute trade, novel environment entry, irreversible action

Tier escalation triggers (Δ factors):

Δ_human: Human presence sensor active in workspace → +1 tier
Δ_trust: Agent trust score below threshold or recent failures → +1 tier
Δ_env: Robot near physical boundary or unstructured environment → +1 tier
Δ_novelty: Action outside validated distribution (novelty detector) → +1 tier

Formal Specification

Request Lifecycle DFA

SINT models every request as a deterministic finite automaton with 12 states:

IDLE → PENDING → POLICY_EVAL → PLANNING → OBSERVING/PREPARING/ACTING → COMMITTING → COMPLETED
                     ↓                              ↓
                ESCALATING                      ROLLEDBACK  (estop, execution failure)
                     ↓
                  FAILED     (approval denied, timeout)

The ACTING state is only reachable via POLICY_EVAL with a valid token. Physical actuation is structurally impossible without a valid capability token.

Tier Assignment Function

Tier(r) = max(BaseTier(r), Δ_human(r), Δ_trust(r), Δ_env(r), Δ_novelty(r))

Formal Invariants

Invariant	Description
I-T1 (Attenuation)	`scope(child_token) ⊆ scope(parent_token)` — delegation can only reduce permissions
I-T2 (Unforgeability)	Capability tokens are Ed25519-signed; valid tokens are computationally unforgeable
I-T3 (Physical Constraint Primacy)	Physical constraints (velocity, force, geofence) in a token cannot be weakened by any downstream layer
I-G1 (No Bypass)	Physical actuation is only reachable from the ACTING DFA state, which is only reachable via POLICY_EVAL
I-G2 (E-stop Universality)	The `estop` event transitions any non-terminal state to ROLLEDBACK unconditionally
I-G3 (Ledger Primacy)	COMMITTING → COMPLETED requires `ledger_committed`; no action completes without a ledger record

Benchmark Results

PolicyGateway latency (measured on M3 MacBook Pro, pnpm run bench):

Tier	p50	p99
T0 (OBSERVE)	~1ms	~3ms
T1 (PREPARE)	~1ms	~3ms
T2 (ACT)	~1ms	~3ms
T3 (COMMIT)	~1ms	~3ms

The gateway adds sub-3ms overhead at p99 for all tiers. Run benchmarks: pnpm run bench.

Key Concepts

Capability Tokens

Ed25519-signed capability tokens — the only authorization primitive. Unlike RBAC (ambient authority to principals), OCap requires explicit token presentation for every operation.

Token fields:

Resource scoping — what the agent can access (ros2:///cmd_vel, mcp://filesystem/*, a2a://agents.example.com/*)
Action restriction — what operations are allowed (publish, call, subscribe, a2a.send)
Physical constraints — max velocity (m/s), max force (N), geofence polygon, time window, rate limit
Delegation chains — max 3 hops, attenuation only (invariant I-T1)
Revocation — instant invalidation via revocation store (ConsentPass endpoint)
W3C DID identity — did:key:z6Mk... format for agent portability

Evidence Ledger

Every policy decision is recorded in a SHA-256 hash-chained append-only log. Chain integrity: ℓ_k.previousHash = SHA256(canonical(ℓ_{k-1})). A gap or hash mismatch constitutes tamper evidence.

Retention policy:

Tier	Retention
T0 (OBSERVE)	30 days
T1 (PREPARE)	90 days
T2 (ACT)	180 days
T3 (COMMIT)	365 days (indefinite if legal hold)

CSML: Composite Safety-Model Latency

A deployment metric that fuses behavioral and physical safety dimensions:

CSML(m, p, t) = α·AR_m + β·BP_m + γ·SV_m - δ·CR_m + ε·𝟙[ledger_intact(t)]

CSML above a deployment threshold θ automatically escalates all subsequent requests from that model backend to the next tier.

Compliance Mapping

IEC 62443 FR1–FR7

FR	Title	SINT Mechanism
FR1	Identification & Authentication	SintCapabilityToken with Ed25519 agent identity; W3C DID portability
FR2	Use Control	Four-tier Approval Gate; `maxRepetitions` constraint; per-resource action allowlists
FR3	System Integrity	SHA-256 hash-chained Evidence Ledger; TEE ProofReceipt for T2/T3
FR4	Data Confidentiality	Zenoh TLS transport; capability scope prevents sensor access without explicit token
FR5	Restricted Data Flow	Policy Gateway allowlists; `geofence` constraint; SINT Bridge per-topic DFA
FR6	Timely Response	`safety.estop.triggered` event; E-stop universality invariant I-G2
FR7	Resource Availability	Per-token rate limiting; `maxRepetitions`; budget enforcement in capsule sandbox

EU AI Act Article 13

Requirement	SINT Approach
Logging and traceability	SHA-256 hash-chained Evidence Ledger — tamper detection is cryptographic
Human oversight	Dynamic Consent + T3 approval gate — T3 actions cannot execute without recorded human approval
Risk management	Tier escalation based on real-time physical context (Δ_human, Δ_env, Δ_novelty)

API Endpoints

Method	Endpoint	Description
`GET`	`/.well-known/sint.json`	Public protocol discovery (version, bridges, profiles, schemas)
`GET`	`/v1/health`	Health check
`POST`	`/v1/intercept`	Evaluate a single request
`POST`	`/v1/intercept/batch`	Evaluate multiple requests (207 Multi-Status)
`POST`	`/v1/tokens`	Issue a capability token
`POST`	`/v1/tokens/delegate`	Delegate (attenuate) a token
`POST`	`/v1/tokens/revoke`	Revoke a token
`GET`	`/v1/ledger`	Query audit ledger events
`GET`	`/v1/approvals/pending`	List pending approval requests
`POST`	`/v1/approvals/:id/resolve`	Approve or deny a request (M-of-N quorum)
`GET`	`/v1/approvals/events`	SSE stream for real-time approval events
`POST`	`/v1/a2a`	JSON-RPC 2.0 A2A protocol endpoint
`GET`	`/v1/metrics`	Prometheus metrics
`GET`	`/v1/openapi.json`	OpenAPI surface for gateway integration

Development Phases

Phase	Description	Tests
Phase 1 (complete)	Security Wedge — capability tokens, PolicyGateway, EvidenceLedger	425
Phase 2 (complete)	Engine Core — bridge-mcp, bridge-ros2, engine packages, persistence, gateway-server	+221 (646)
Phase 3 (complete)	Economy Bridge — @sint/bridge-economy with port/adapter pattern, EconomyPlugin	+91 (737)
Phase 4 (complete)	Standards Alignment — A2A bridge, rate limiting, M-of-N quorum, W3C DID identity	+78 (815)
Phase 5 (complete)	Protocol Surface v0.2 — discovery/OpenAPI/schema endpoints, industrial profiles	shipped
Phase 6 (complete)	Engine layer — System1/2 engines, HAL, capsule sandbox, Avatar/CSML, reference capsules	shipped

Deployment

Railway (Recommended)

brew install railway
railway login
./scripts/railway-setup.sh
railway variables --set SINT_STORE=postgres SINT_CACHE=redis SINT_API_KEY=$(openssl rand -hex 32)
railway up

Docker Compose

docker-compose up
# Gateway:   http://localhost:3100
# Dashboard: http://localhost:3201
# Postgres:  localhost:5432
# Redis:     localhost:6379

Tech Stack

Runtime: Node.js 22+
Language: TypeScript 5.7 (strict mode)
Monorepo: pnpm workspaces + Turborepo
HTTP: Hono
Validation: Zod
Crypto: @noble/ed25519, @noble/hashes (audited, zero-dependency)
MCP SDK: @modelcontextprotocol/sdk
Dashboard: React 19, Vite 6
Testing: Vitest (815 passing tests)
Infra: Docker, PostgreSQL 16+, Redis 7, GitHub Actions CI, Railway

Docs & Artifacts

Protocol spec: docs/SINT_v0.2_SPEC.md
SIP governance: docs/SIPS.md
Release notes: docs/RELEASE_NOTES_v0.2.md
Conformance matrix: docs/CONFORMANCE_CERTIFICATION_MATRIX_v0.2.md
Deployment profiles: docs/profiles/
Examples: examples/ (hello-world, warehouse-amr, industrial-cell)
Multi-language SDKs: sdks/ (TypeScript, Python, Go)
Benchmark report: docs/reports/industrial-benchmark-report.md

Design Principles

Single choke point — Every agent action flows through PolicyGateway.intercept(); no bridge adapter makes authorization decisions independently
Result<T, E> over exceptions — All fallible operations return discriminated unions, never throw
Attenuation only — Delegated tokens can only reduce permissions, never escalate (I-T1)
Append-only audit — The evidence ledger is INSERT-only with SHA-256 hash chain integrity (I-G3)
Physical safety first — Velocity, force, and geofence constraints live in the token, not in external config
Interface-first persistence — Storage adapters implement clean interfaces; swap in-memory for Postgres/Redis
Fail-open on infrastructure — Economy/rate-limit infrastructure failures do not block the safety path
E-stop universality — The hardware E-stop bypasses all token checks and is unconditional (I-G2)

References

ROSClaw: Empirical safety analysis of LLM-controlled physical AI — arXiv:2603.26997 (IROS 2026)
MCP Security Analysis: Architectural vulnerabilities in the Model Context Protocol — arXiv:2601.17549
IEC 62443: Industrial automation and control systems cybersecurity standard
EU AI Act Article 13: Transparency requirements for AI systems
NIST AI RMF: AI Risk Management Framework
W3C DID Core: Decentralized Identifiers specification

License

Apache-2.0

Name		Name	Last commit message	Last commit date
Latest commit History 41 Commits
.github		.github
apps		apps
capsules		capsules
docs		docs
examples		examples
packages		packages
scripts		scripts
sdks		sdks
.dockerignore		.dockerignore
.env.example		.env.example
.gitignore		.gitignore
AGENTS.md		AGENTS.md
CLAUDE.md		CLAUDE.md
CODE_OF_CONDUCT.md		CODE_OF_CONDUCT.md
CONTRIBUTING.md		CONTRIBUTING.md
LICENSE		LICENSE
README.md		README.md
SECURITY.md		SECURITY.md
WHITEPAPER.md		WHITEPAPER.md
docker-compose.yml		docker-compose.yml
package.json		package.json
pnpm-lock.yaml		pnpm-lock.yaml
pnpm-workspace.yaml		pnpm-workspace.yaml
sint-mcp.config.example.json		sint-mcp.config.example.json
tsconfig.base.json		tsconfig.base.json
turbo.json		turbo.json

Folders and files

Latest commit

History

Repository files navigation

SINT Protocol

Why SINT?

Quick Start

Start the Gateway Server

Run a Single Package

For AI Agents

Architecture

APS vs SINT Primitives

Packages

Gate (Security Core)

Bridges (11 bridges)

Engine (AI Execution Layer)

Reference Capsules

Persistence

Apps & SDKs

Approval Tiers

Formal Specification

Request Lifecycle DFA

Tier Assignment Function

Formal Invariants

Benchmark Results

Key Concepts

Capability Tokens

Evidence Ledger

CSML: Composite Safety-Model Latency

Compliance Mapping

IEC 62443 FR1–FR7

EU AI Act Article 13

API Endpoints

Development Phases

Deployment

Railway (Recommended)

Docker Compose

Tech Stack

Docs & Artifacts

Design Principles

References

License

About

Topics

Resources

License

Code of conduct

Contributing

Security policy

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages