CSGHub Release Notes v2026.10
1 AI Gateway Optimization
Comprehensively enhanced gateway high-concurrency stability, security governance, and standard alignment.
- Prompt Injection Defense & Sensitive Data Leak Prevention: Added Prompt injection detection and end-to-end sensitive data filtering to prevent confidential information leakage in requests, responses, and logs.
- Whitelist Access Control & Model Permission Isolation: Added IP whitelist access control and supported restricting callable model scopes based on API Keys.
- High Concurrency Stability & Connection Pooling: Integrated connection pooling into health checks and optimized concurrent upstream probing, preventing 429 surges from causing false node eviction and circuit breaker deadlocks.
- OpenAI Standard Alignment & Token SLO Metrics: Standardized error status codes and error messages to align with OpenAI formats, and launched the Token production Service Level Objective (SLO) metric monitoring system.
2 Heterogeneous Inference Engine Upgrades & Frontier Model Support
Optimized end-to-end LLM generation latency while expanding frontier reasoning model and domestic compute ecosystems.
- Core Inference Runtime Upgrade: Fully upgraded underlying inference runtimes (vLLM to v0.31.0, SGLang to v0.5.21), significantly improving throughput and long-context stability.
- Speculative Decoding & Auto Parameter Tuning: Supported Speculative Decoding and Multi-Token Prediction (MTP) to reduce generation latency; added auto parameter tuning APIs to recommend optimal deployment configurations.
- New Decision Model & Domestic Compute Adaptation: Added support for the Laya fast-thinking decision model inference engine; completed end-to-end compatibility with MetaX GPU domestic compute platforms.
3 MCP Gateway Statelessness & Agent Infrastructure (EE)
- Stateless MCP Gateway Architecture: Completed the stateless architectural refactoring of the Model Context Protocol (MCP) gateway, enabling horizontal elastic scaling for microservices.
- Self-Hosted MCP Server Integration: Enabled users to register and mount self-hosted MCP Servers onto the platform gateway, rapidly extending agent toolchains.
- Agent Orchestrator Runtime Scheduling: Implemented isolated runtime orchestration for shared Agents, ensuring every user enjoys clean, independent execution sessions and workspaces.
4 Commercial Token Subscription & Financial Lifecycle
Streamlined the commercial lifecycle across template configuration, subscription purchase, automatic credit deduction, and usage reconciliation.
- Tiered Token Subscription (Token Plan): Supported personal and organizational subscription plans with automatic Credit balance deduction, alongside configurable admin subscription templates and accounting reports.
- Transparent Billing & API Quota Governance: Supported transparent subscription usage views in user/org billing centers; Enterprise Edition (EE) added API Key quota allocations and temporary enable/disable capabilities.
5 Multi-Level Organizations & Fine-Grained Authorization (EE)
Refined enterprise multi-tenant management architectures to ensure seamless permission governance across large teams.
- Batch Import of External Organizational Hierarchies: Supported one-click synchronization and import of existing corporate HR/LDAP multi-level organizational trees, minimizing tenant initialization costs.
- Hierarchical Administrator Role Delegation: Supported multi-tiered administrator role authorization, restricting admins at different tiers to managing assets and members within their respective sub-organizations.
6 Security & Compliance Hardening
Hardened platform identity authentication chains to eliminate privilege escalation and data exposure vulnerabilities.
- Authentication Redirect Hardening & Credential Protection: Fixed Open Redirect and JWT exposure risks caused by trusting redirection URLs during login callback flows; unified Portico OIDC integration and bound OIDC Issuer to prevent collision attacks.
- API Redaction & Strict CORS Policies: Fixed email address exposure in public endpoints, tightened global CORS policies across services, and sanitized system logs to prevent recording sensitive tokens and keys.
History Versions
Click Release List to view all historical versions.