Skip to main content
This is the canonical whitepaper as of April 2026. The full academic version is maintained in the repo at github.com/sint-ai/sint-protocol/blob/main/WHITEPAPER.md. The arXiv preprint is under submission (cs.CR primary, cs.RO / cs.AI / cs.MA cross-list). The March 2025 “SINT Agentic Platform” whitepaper is archived at /archive/2025-whitepaper.

Read in GitHub

The canonical Markdown version with full bibliography and citation metadata.

arXiv preprint

Academic version with LaTeX typesetting. Under submission as of April 2026.

​
Abstract

Foundation-model AI agents now perceive, reason, and physically actuate. Coinbase’s Agentic Wallets (February 2026) gave them money; ERC-8004 gave them portable on-chain identity; OWASP’s Top 10 for Agentic Applications (December 2025) catalogued how they fail; the EU AI Act’s high-risk provisions take effect August 2, 2026. None of these infrastructures addresses the full stack of a physical-AI agent — the combined cognition, identity, authorization, actuation, and audit surface that spans software and the physical world. This paper presents SINT (Secure Infrastructure for Non-human Transactions), a protocol that formalizes physical-AI agents as governed entities possessing typed identity, graduated capability rights, and an immutable civic record.

​
Contributions

SINT contributes:
  1. A formal SintCapabilityToken schema with a physical-constraints block, cryptographically bound to a W3C DID and a TEE attestation.
  2. A twelve-state deterministic finite automaton governing the full agent authorization lifecycle, mapped to ROS 2 action primitives.
  3. A four-tier human oversight model with a provable max-escalation function driven by human-presence, environmental, novelty, and composite-safety signals.
  4. A STRIDE-extended threat model — STRIDE+B — whose seventh class formalizes behavioral non-determinism across foundation-model backends.
  5. A seven-benchmark evaluation framework with per-tier latency budgets and a 157-test conformance suite.
  6. Compliance mappings to IEC 62443, EU AI Act Article 13, and NIST AI RMF.
We situate SINT as the security substrate that ROSClaw’s executive-layer contract C = ⟨A, O, V, L⟩ explicitly scopes out, complementary to SAGA’s inter-agent trust layer, and grounded in the empirical pattern that identical commands issued across frontier model backends produce a 3.4× spread in prompt-level out-of-policy attempt rates.

​
Core sections

Formal specification

The normative SintCapabilityToken schema, wire messages, and DFA state transitions.

Tier model

T0–T3 with the max-escalation function and Δ factors.

CSML

The Composite Safety-Model-Ledger score, calibrated against ROSClaw data.

Threat model

STRIDE+B with worked attack chains for UniPwn and MCP goal hijack.

Invariants

Safety, liveness, non-repudiation, tamper-evidence proofs.

Compliance

IEC 62443 / EU AI Act / NIST AI RMF mappings.

​
Download

Markdown

Raw source on GitHub.

PDF

Released alongside each protocol version tag.

BibTeX

For academic citation.

​
Citing SINT

@techreport{sint_protocol_2026,
  title        = {{SINT}: A Formal Security Protocol for Governed Physical-{AI} Agents},
  author       = {Pashkov, Illia and {SINT Labs}},
  institution  = {SINT Labs},
  year         = {2026},
  month        = {April},
  note         = {v0.2. Open specification (CC BY 4.0), reference implementation (Apache-2.0).},
  url          = {https://github.com/sint-ai/sint-protocol}
}
When the arXiv preprint is public, replace the citation with the arXiv BibTeX entry (same key). Watch the facts page for updates.
Implementing SINT from the paper? Also read the protocol spec and the conformance suite. The spec is the normative artifact; the whitepaper is the explanatory one.