Independent Source For China's Tech Innovation Since 2000

How Did China Mobile's AI Agent Score on a Brutal Global Cybersecurity Benchmark?

An artificial intelligence cybersecurity agent jointly developed by China Mobile has achieved a 96.35% vulnerability reproduction success rate on a prominent international benchmark, ranking first globally among telecommunications operators and state-owned enterprises, according to state media reports.

The offensive and defensive AI system, named "TieYing," or Iron Eagle, was developed by China Mobile Group's Network Department in collaboration with the Network Management Center of its provincial subsidiary, Zhejiang Mobile, the Qianjiang Evening News boasted Wednesday.

The result places China Mobile's AI-native security technology among top global performers at a time when industry analysts say digital transformation and cloud adoption are rapidly expanding cyberattack surfaces. As automated attacks and vulnerability iterations accelerate, cybersecurity experts note that digital defense is shifting from human-versus-human confrontations to automated battles between opposing AI systems.

TieYing's performance was recorded on the latest leaderboard for CyberGym, an authoritative AI offensive and defensive evaluation framework created by researchers at the University of California, Berkeley, and recognized by major international academic conferences.

Designed to test the real-world security capabilities of autonomous AI agents, CyberGym draws on a large database of historical vulnerabilities from open-source software projects to simulate authentic cyber conflict scenarios. The benchmark evaluates participants on code comprehension, logical reasoning and vulnerability assessment, attracting entries from major global technology firms including Microsoft, Google, OpenAI and Alibaba.

Developers said TieYing's score stems from a specialized multi-agent collaborative architecture built to overcome traditional bottlenecks in vulnerability analysis, such as lengthy investigation cycles, complex code logic and weak feedback loops.

At the core of the platform is a global scheduling system driven by a Meta-Agent, which breaks down complex security tasks into granular subtasks and executes them in parallel across multiple threads to speed up analysis.

To prevent blind spots and false positives common in single-path reasoning, the system cross-validates findings across five specialized analytical modules: taint tracking, signature recognition, business logic inspection, supply chain risk evaluation and vulnerability pattern hunting.

Supported by a hierarchical vulnerability knowledge base, dynamic atomic tool invocation and full-process state memory, the agent can sustain iterative reasoning over extended task cycles, retaining valid investigative leads while discarding dead ends to autonomously discover and reproduce software flaws.

Unlike general-purpose large language models that focus on scaling parameter counts, China Mobile engineered TieYing specifically for practical security production, emphasizing accurate code reading, precise flaw localization and reliable exploit verification.

China Mobile plans to transition TieYing from benchmark testing into live network security operations. Upcoming deployments will focus on automated code auditing, pre-launch risk inspections, legacy vulnerability reviews and routine threat assessments to achieve early detection and closed-loop remediation, reducing reliance on manual security operations.

Other China Tech Buzz: