Skip to content

Request: Include SINT Protocol in AI Security Solutions Landscape for Agentic AI #11

@pshkv

Description

@pshkv

Requesting inclusion of SINT Protocol in the AI Security Solutions Landscape for Agentic AI.

Tool: SINT Protocol
Type: Open-source runtime authorization framework for physical AI agents
License: Apache 2.0
Repository: https://github.com/sint-ai/sint-protocol

What it does:
SINT Protocol is a runtime safety shield that interposes cryptographic authorization at every LLM-agent–actuator boundary. Every agent action (tool call, robot command, code execution) passes through a single Policy Gateway that enforces capability tokens, graduated human oversight, behavioral drift detection, and physical constraint envelopes before hardware execution.

Coverage:

  • OWASP Agentic Top 10: 10/10 coverage (ASI01–ASI10)
  • ASI01 Goal Hijack: 5-layer heuristic detection (role override, semantic escalation, exfiltration probes, cross-agent injection, prompt injection in tool args)
  • ASI02 Tool Misuse: Tier-based authorization (T0 observe → T3 commit), forbidden tool combo detection
  • ASI03 Identity Abuse: Ed25519 capability tokens with subject binding, delegation depth enforcement (max 3 hops), cascade revocation
  • ASI04 Supply Chain: Model fingerprint hash + model ID allowlist validation at runtime
  • ASI05 Code Execution: Shell/exec/eval tools auto-classified T3_COMMIT (human sign-off required)
  • ASI06 Memory Poisoning: Replay detection, privilege claim detection, history overflow, cross-session continuity injection, credential read funnel (≥3 secret reads + write), action velocity loops
  • ASI07 Inter-Agent: SwarmCoordinator with collective constraints (max concurrent actors, kinetic energy ceiling, minimum inter-agent distance)
  • ASI08 Cascade: CircuitBreakerPlugin — N consecutive denials auto-trip, CSML anomalous-persona auto-trip, HALF_OPEN probe recovery
  • ASI09 Trust Exploitation: ProactiveEscalationEngine — CSML behavioral drift monitoring across multi-turn windows
  • ASI10 Rogue Agent: EU AI Act Art. 14(4)(e) compliant emergency stop with manual trip + auto-recovery

Key differentiator: Only framework with physical constraint enforcement (velocity, force, geofence) in cryptographic tokens. Designed for robots, drones, industrial actuators — not just digital/software agents.

Bridge adapters: ROS 2, MAVLink v2, MCP, Google A2A, MQTT/CoAP, OPC-UA, Open-RMF, Sparkplug B, gRPC (12 total)

Stats:

  • 42 packages, 1,710+ tests
  • Gateway latency: p99 < 5ms steady-state (within 100Hz ROS 2 control loop budget)
  • TypeScript monorepo, zero external runtime dependencies for crypto (@noble/ed25519)

Lifecycle stage: Runtime Authorization / Policy Enforcement / Audit

Academic submission: 4-page short paper submitted to SPAI 2026 (1st IJCAI Workshop on Safe Physical AI, Bremen, August 2026)

Comparison with existing landscape entries:

  • vs Microsoft AGT: AGT targets digital/software agents; SINT targets physical AI with hardware constraint enforcement
  • vs ATR: ATR provides detection rules (static analysis); SINT provides runtime enforcement (policy gateway). Complementary.

Happy to provide additional information or a demo walkthrough.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions