Introduction

As AI agents transition from "point pilots" to "large-scale production environments," AI agent security governance has become a must-answer question for enterprises pursuing digital transformation. In IDC's China Agent Threat Detection Technology Assessment, 2026 (Doc# CHC53616626, April 2026), Alibaba Cloud stood out among more than 20 participating vendors, drawing on its deep expertise across asset management, posture management, data security, and technological innovation. Alibaba Cloud Agent Security Center earned 4 perfect scores (5/5) and achieved high scores in three additional dimensions: vulnerability detection, runtime detection, and MCP security detection. These results affirm Alibaba Cloud's product and technology leadership in the field of AI agent security.

Leading Capabilities: The Technology Behind Four Perfect Scores

IDC's assessment focused on the risks that agents introduce in enterprise environments — data leakage, unauthorized operations, and business disruption — with particular attention to vendors' capabilities for threat detection and governance across the full agent lifecycle. Alibaba Cloud's strong performance across all seven dimensions stems from its departure from the traditional single-point defense approach — such as relying solely on LLM security gateways or safety guardrails — and its construction of a defense system built on the philosophies of "agent-native security" and "defending AI with AI."

Comprehensive visibility into agent assets: eliminating "shadow agent" blind spots

In the "Agent Asset Management" dimension, Agent Security Center builds a precise "agent asset map" that automatically discovers and continuously monitors AI applications across cloud environments. The product covers 5 major cloud platforms and can identify more than 190 AI components. By generating a visualized "agent relationship graph," it deeply correlates isolated agent applications with their underlying models, Skills, knowledge bases (RAG), tools, and identity credentials — resolving the security blind spots caused by black-box assets in traditional approaches.

Beyond cloud asset coverage, Alibaba Cloud also extends shadow agent discovery and governance to office security scenarios through its Alibaba Cloud SASE 2.0 product.

Closed-loop posture management: dynamic governance of the supply chain and identities

In the "Agent Security Posture Management" dimension, targeting risks such as agent supply chain poisoning and misconfiguration, Alibaba Cloud introduced Agent-SPM (Agent Security Posture Management) and Agent-BOM (Agent Bill of Materials) capabilities, achieving automated compliance checks across the entire pipeline — from image building in the development stage, through deployment, to agent runtime. For identity governance, Agent ID Guard delivers fine-grained access control for machine-to-machine (M2M) identities, using temporary credentials and short-lived permission mechanisms to effectively curb credential abuse and privilege escalation risks.

Defense in depth for data security: end-to-end control from source to flow

In the "Agent Skill and Data Security Detection" dimension, facing the complex scenarios agents encounter when processing sensitive data, Alibaba Cloud Data Security Center (DSC) adopts a hybrid detection architecture combining "rule engine + semantic understanding + knowledge reasoning." This architecture not only identifies structured sensitive information (such as ID card numbers and bank card numbers) at millisecond speed, but also leverages the semantic analysis capabilities of the Qwen LLM to precisely detect implicit sensitive content in unstructured documents. More importantly, Alibaba Cloud has built an end-to-end data lineage management platform, achieving full lifecycle data tracking from knowledge bases and RAG systems to Skills components — ensuring that data remains traceable and controllable throughout its flow.

Innovation and frontier technology convergence: AI vs. AI

In the "Innovation" dimension, Alibaba Cloud earned a perfect score for its "defending AI with AI" technical approach, reflecting the evolution of security capabilities from "resource stacking" to "agent-native."

Facing emerging threats such as prompt injection and jailbreak attacks, Alibaba Cloud performed deep security fine-tuning on the Qwen foundation model and built a new generation of proactive detection systems — an automated AI Red Teaming system. The built-in "attack agent" simulates real attacker strategies, automatically generates adversarial samples, and conducts simulated attack testing against models — representing the evolution from passive "protective resources" into growing, collaborative, and self-evolving "digital immune agents." In addition, Alibaba Cloud has distilled its core detection capabilities into the open-source model Qwen3-Guard, demonstrating an open posture toward ecosystem co-building and technology contribution.

Runtime Detection and MCP Security: A Dynamic Defense System for the AI-Native Era

Beyond the perfect-score dimensions above, Alibaba Cloud demonstrated robust and outstanding defensive capabilities in complex attack scenarios across three additional dimensions: "Vulnerability Detection & Management," "Runtime Detection," and "MCP Security Detection."

● Vulnerability Detection & Management: Agent Security Center fuses traditional CVE scanning with AI-native risk identification, analyzes the supply chain in real time, and provides precise remediation guidance — balancing security hardening with AI environment stability.

● Runtime Detection: AI Safety Guardrails 2.0 performs behavioral correlation analysis based on the complete execution chain (request → orchestration → inference → tool invocation → response), dynamically intercepting jailbreaks, tool poisoning, and anomalous operations across cloud, endpoint, and multi-agent collaboration scenarios.

● MCP Security Detection: Through Agent ID Guard, Alibaba Cloud enforces identity authentication and least-privilege verification for MCP (Model Context Protocol) invocations, combined with dynamic credentials hosted in KMS. Even if an MCP server is compromised, attackers cannot obtain long-lived valid credentials — fortifying the front line of agent interactions.

From Technical Leadership to Real-World Effectiveness

Alibaba Cloud Agent Security Center's strong results in this IDC assessment represent the culmination of years of native technology investment in AI security infrastructure. From full-stack protection at the infrastructure layer, to content safety at the model layer, to behavior control at the application layer, this system has been successfully deployed among enterprise customers across finance, new retail, logistics, and other industries — effectively addressing the new security challenges brought by AI agents.

After deploying Alibaba Cloud Agent Security Center, a leading investment community successfully remediated OpenClaw configuration risks and third-party Skills supply chain vulnerabilities. Through a unified onboarding solution, the enterprise automatically discovered dozens of shadow agents and established an asset inventory. Full-chain prompt protection intercepts hundreds of attacks daily on average, sensitive information leakage incidents have dropped to zero, and malicious Skill executions are effectively isolated in a sandbox.