// Security Whitepaper

Securing agent workflows in the LLM router era.

How XinoAPI reduces supply-chain risks for developers using DeepSeek, Qwen, GLM, Kimi, and MiniMax through an intermediary gateway.

Date: June 2026 Subject: Chinese LLM inference export Reference: arXiv:2604.08407v1 Status: Implementation ready + roadmap

Executive Summary

Multi-model AI stacks increasingly depend on API routers to reach high-performance and cost-effective models. That router layer is now part of the software supply chain.

XinoAPI is designed as a security-aware AI gateway: OpenAI-compatible access to Chinese frontier models, metadata-based billing, response scanning, gateway signing, and an open-source privacy SDK for local redaction. The goal is to reduce router risk without forcing teams to give up model choice.

The Intermediary Problem

Recent research on malicious LLM API intermediaries describes two practical attack classes for agents that read, write, and execute code.

AC-1

Payload Injection

A compromised router or upstream path can alter an LLM response to include malicious shell commands, unsafe imports, or tool instructions that an autonomous coding agent may execute.

AC-2

Secret Exfiltration

Tool arguments, repository context, prompts, API keys, database credentials, or cloud configuration may leak through intermediary logs or model-bound payloads.

Important scope note: XinoAPI reduces and monitors these risks. No intermediary can honestly claim to eliminate all model, user, upstream, or agent-execution risk.

Layered Defense System

The security posture combines local controls, gateway inspection, signed delivery, and enterprise isolation options.

Layer 1

Local-first privacy SDK

Client-side redaction can mask PII, credentials, and high-risk secrets before prompts leave your environment. This keeps sensitive material out of router and upstream logs.

Layer 2

Gateway payload shield

Streaming inspection scans model output for high-risk response patterns such as unsafe shell execution, suspicious Python imports, and tool-call abuse indicators.

Layer 3

Response signing

Gateway signatures can help clients verify that the response delivered by XinoAPI was not modified after it left the gateway. Upstream-native proof requires upstream support.

Available Now vs Roadmap

Capability Status Security Value
Metadata-only billing logs Available Supports usage accounting without routine prompt/response body retention.
Security service health Available Gateway signing, scanning, and audit services are exposed through the XinoAPI security layer.
Privacy SDK Available / evolving Local redaction and client-side controls for sensitive agent context.
Anthropic-compatible bridge Roadmap Claude Code style workflows through /anthropic/v1/messages.
Trusted execution environments Enterprise roadmap Optional isolated execution environments such as AWS Nitro Enclaves for sensitive deployments.
Third-party security audit Planned Independent validation of no-content-retention claims and gateway controls.

Technical Comparison

XinoAPI is positioned for teams that care about reliability, auditability, and Chinese-model coverage rather than the lowest possible token price.

Feature Generic Routers XinoAPI
Response sanitization Usually raw forwarding Active scanning for high-risk response patterns
Data privacy model Trust-based router handling SDK-based local masking plus metadata billing logs
Agent workflows Generic chat routing Codex, OpenClaw, Cline/Roo focus; Claude bridge roadmap
China model coverage Varies by marketplace DeepSeek, Qwen, GLM, Kimi, MiniMax specialization
Commercial support Self-serve first Stripe, $20 minimum top-up, invoice/SLA path for teams

Compliance Posture

Privacy

GDPR / CCPA alignment

XinoAPI is designed to support enterprise privacy workflows through local redaction, metadata-based usage records, and published terms and privacy policy pages.

Auditability

Operational evidence

Billing events, model usage, gateway health, and security-layer metadata provide audit-ready operational visibility without treating prompt content as routine billing data.

Conclusion: Security in the LLM supply chain is no longer optional. XinoAPI gives international developers a practical path to use Chinese frontier models while reducing agent-specific intermediary risk.