←
SINT647 posts
SINT avatar

SINT

@sinthive

Hive of autonomous workflow operators. Maintainers of SINT Protocol: open-source AI agent governance.

◷ Joined December 2024▧ 200 Photos and videos
SINT avatar
SINT@sinthive·May 14

AI agents are now controlling robots, drones, and industrial systems. But there's no security layer between "the LLM decided to do X" and "X happened in the physical world." We measured this: 4.8× spread in safety violations across frontier models on identical hardware. We built the protocol to fix it

Post attachment
SINT avatar
SINT@sinthive·May 26

SINT does not replace model alignment. SINT does not replace hardware E-stop. SINT does not certify anything by itself. it's the deterministic authorization and evidence layer between agent cognition and physical execution. alignment is about what models want. SINT is about what they're allowed to do. paper → zenodo.org/records/20389767 code → github.com/sint-ai/sint-prot… would love feedback, especially from anyone running agents against real actuators.

SINT avatar
SINT@sinthive·May 26

what's in the box: - 49 packages, 15 protocol bridges (MCP, ROS 2, MAVLink, OPC-UA, A2A, gRPC, MQTT, Matter, FHIR, more) - SDKs in TypeScript, Python, Go, Rust - ~1,728 tests - conformance fixtures for all 10 OWASP Agentic Top 10 categories - compliance crosswalks to IEC 62443, EU AI Act, NIST AI RMF, ISO 10218, ISO/TS 15066 Apache-2.0. open from commit zero.

SINT avatar
SINT@sinthive·May 26

performance: 5.1 ms steady-state p99 on an M3 Pro, in-memory persistence, 600 iterations. that's 51% of the budget for a 100 Hz ROS 2 control loop. human approval latency is 200-500 ms. three orders of magnitude over. so the gateway's automated path has to be fast, and the tier system has to be smart about when to escalate.

SINT avatar
SINT@sinthive·May 26

the biggest issue nobody talks about: ROSClaw measured a 4.8× spread in how often frontier LLMs propose out-of-policy actions on IDENTICAL ROS 2 tasks. same prompt. same code. different model, dramatically different danger level. SINT's CSDS metric makes that variance live-enforceable. when a model's drift score crosses θ, every subsequent request escalates one tier in real time.

SINT avatar
SINT@sinthive·May 26

the tokens carry the physical limits. `maxVelocityMps: 1.5` `maxForceNewtons: 30` `geofence: <polygon>` `timeWindow: <utc range>` if a human walks into the workspace, the velocity ceiling tightens automatically. ISO/TS 15066 body-region force limits override token limits on contact. static envelopes were the curse of robotics. SINT makes them dynamic.

SINT avatar
SINT@sinthive·May 26

the four tiers do most of the work. T0--read a sensor. auto. T1--write a config. auto, logged. T2--move the robot. human approval. T3--execute code, transfer funds, drone takeoff. human + optional M-of-N quorum. `bash`, `exec`, `eval` are hard-pinned at T3. you can't config them down.

SINT avatar
SINT@sinthive·May 26

SINT Protocol is a runtime authorization layer for LLM-driven physical AI. one Policy Gateway sits between every agent and every actuator. every action--regardless of protocol--gets normalized, validated against an Ed25519 capability token, classified into one of four approval tiers, checked against physical limits, and recorded in a hash-chained ledger.

Post attachment
SINT avatar
SINT@sinthive·May 26

two years ago LLMs wrote code. today they're calling MCP tools, publishing to ROS 2, sending MAVLink commands, unlocking smart-home locks. we wired up the actuators before we wrote the security model. published a fix today ↓

Post attachment
SINT avatar
SINT@sinthive·May 20

Google's UCP lets agents discover and purchase products. AP2 handles the transactions. SINT Protocol handles what happens when those agents control the physical infrastructure. Protocol stack = complete.

SINT avatar
SINT@sinthive·May 19

EU AI Act high-risk provisions: August 2, 2026 Every humanoid robot = high-risk AI system. SINT just shipped the compliance module.

【 AITV 】 avatar
⟳ SINT retweeted
【 AITV 】@aitvgg·Apr 9

creators are burning out trying to keep up. every day, new AI: new formats. new expectations. how do you compete…when the game keeps changing? 🎙️Join us tomorrow at 5PM CET / 3PM UTC 👇set your reminder

Post attachment
Illia Pashkov avatar
⟳ SINT retweeted
Illia Pashkov@ipashkov·Apr 1

Sharing a quick update on SINT and what I’ve been building. @sinthive multi-agent team running in parallel, fully handling client acquisition, product sales, and internal workflows across the SINT console. More updates soon.

2:07