ShouYuan LLM Security Guardrail: Smart Safety Locks for OpenClaw Agents 为OpenClaw们装上智能安全锁，长亭发布”守元”大模型安全围栏

Originally published on Chaitin Technology’s WeChat account. Read original

OpenClaw went viral — the era of mass AI agent adoption is here. But security concerns are more urgent than ever.

According to CNNVD, as of March 9, OpenClaw has 82 known vulnerabilities, with 33 rated critical or high severity. The security landscape is alarming.

As a leader in intelligent security, Chaitin Technology officially launches “ShouYuan” (守元) LLM Security Guardrail — featuring innovative streaming async detection, multi-model hybrid detection, data flywheel self-evolution, and agent risk cognition technologies.

Four AI Risks, One Solution

Content Safety: Auto-filters hate speech, violence, pornography, and value-misaligned content.

Model Jailbreak Attacks: Prevents carefully crafted prompts from bypassing safety limits.

Data Leakage: Real-time detection of sensitive information (IDs, bank accounts, trade secrets) in inputs/outputs.

Unsafe Execution: Prevents agents from performing unauthorized, non-compliant, or destructive operations.

Streaming Async Detection: Performance Meets Security

Traditional approaches use “generate → full scan → return,” causing latency. ShouYuan innovates with streaming async detection: while the model generates tokens, a parallel copy streams to the security engine for real-time risk assessment. Risks trigger immediate alerts or interceptions — zero perceptible delay.

Multi-Model Hybrid Detection

ShouYuan 1.0 employs a lightweight dynamic multi-model recognition system:

• Text models: 8B, 4B, and 0.6B variants, aligned to national standards (5 major categories, 31 subcategories), supporting Chinese, English, and 50+ dialects
• Small models: RoBERTa-based, trained on 800K proprietary samples with long-sequence understanding
• Multimodal: VL and Whisper models covering images, audio, and video risks

Models are orchestrated via DAG graphs to form a collaborative “intelligent defense matrix.”

Data Flywheel Self-Evolution

ShouYuan integrates offense (“spear” — security evaluation) and defense (“shield” — guardrail) through full-pipeline data exchange. By identifying false positives and false negatives in real deployments, the system continuously self-adapts to each customer’s unique scenario.

Agent Risk Cognition

For agents like OpenClaw operating in open environments, risks span the entire workflow — input, reasoning, tool calls, and execution. ShouYuan’s proprietary agent risk framework decomposes risks across 4 key dimensions and defines 10 agent-specific risk categories (task intent hijacking, unsafe tool calls, etc.) for precise full-chain security diagnosis.