appConfigEntityId

bodyType

customContents

entityId

entityUuid

isVideo

articlePageQuery

article

articleSchemaContent

content

hooksArticlePageAdvertisementConfigArticle

hooksArticlePageFloatingButtonConfigArticle

appConfig

filter

json

liveArticle

Vincent Chow

Chinese artificial intelligence systems are nearing levels of “frontier risks” similar to their US counterparts, according to a study.
Recent advances in the capabilities of AI models from DeepSeek and other Chinese firms heightened the possibility that these systems could be misused by bad actors or escape human control, according to Concordia AI, a Beijing-based consultancy focused on AI safety in China.
As AI models grew more powerful, their frontier risks – or their potential to endanger public safety and social stability – have raised alarms among experts about possible catastrophic consequences, including the destruction of humanity.
Concordia AI’s assessment, based on its analysis of 50 leading AI models and shared exclusively with the Post, indicated that Chinese models had joined their US counterparts in pushing the boundaries of such risks.
“We hope our findings can support these companies as they improve the safety of their models,” said Fang Liang, the consultancy’s head of AI safety and governance in China.

DeepSeek’s flagship R1 model, last updated in May, was identified as having the highest risk score for carrying out cyberattacks.
ByteDance’s Doubao Seed 1.6 was the first Chinese model to surpass a human expert in troubleshooting biological protocols, alongside OpenAI’s GPT-5 and xAI’s Grok 4, the report showed. The capability is correlated with potential bioweapons development.
Concordia AI had sent customised reports of its evaluation to several companies, including DeepSeek and Kimi developer Moonshot AI, according to Fang. DeepSeek and Moonshot did not respond to requests for comment on Friday.
In recent months, the Chinese government has expressed concerns about frontier risks.
A technical standards body associated with the Cyberspace Administration of China noted the risk of AI systems escaping human control in a document last month, while the country’s cybersecurity law was amended last week to include provisions on AI safety and incident reporting for the first time.
Wang Weibing, Concordia AI’s safety research senior manager who led the study, said that the risk scores for each model were determined by evaluating their capabilities and safety performance against industry benchmarks across four domains: cyber offence, biological risks, chemical risks and loss of control.
Overall, Concordia AI found that the aggregate risk scores for all models had risen across all domains over the past year, with loss-of-control risks soaring 50 per cent and biological risks jumping 38 per cent.
A key driver of these increases was the proliferation of reasoning models, according to Concordia AI CEO Brian Tse.
These models, which “think” before generating output, prompted US firms like OpenAI and Anthropic to enhance their safety measures. In the study, their respective GPT and Claude models maintained relatively lower risk levels despite improved capabilities.
Concordia AI planned to make its findings publicly available online, a move that would make it the first third-party platform in China dedicated to evaluating frontier AI risks, according to Fang.
Future plans included assessing other risk domains, such as large-scale deception and psychological manipulation risks, he added.


Alibaba teases 10-trillion-parameter model, debuts ‘China’s most powerful’ AI chip

Beyond Phoebe Gates’ Phia: 5 female-owned fashion-tech apps to have on your radar

Ant Group consolidates Alipay operations in big bet on AI ‘agentic commerce’

US, China need nuclear-style AI safeguards, congressman says before Xi-Trump summit

Artificial intelligence (AI)

Chinese AI models comparable to US ones in frontier risks, study finds

Chinese AI models have joined their US counterparts in pushing the boundaries of frontier risks, according to a new report. Photo: Shutterstock

Chinese AI start-up DeepSeek’s R1 model scored the highest in cyber-offence risks in Concordia AI’s assessment. Photo: Zuma Press Wire/TNS

Alibaba’s Qwen returns 22 per cent in 2 weeks, beats DeepSeek, OpenAI in crypto trading showdown

Alibaba Cloud’s Qwen3-Max model generated a 22.32 per cent return on an initial investment of US$10,000 over two weeks. Photo: Shutterstock

AI coding tools built by US firms face scrutiny over Chinese model origins

Two US companies released AI tools with powerful coding capabilities and rapid generation speed. Photo: Shutterstock

Zhipu AI said it believed US company Cognition AI used its latest flagship GLM-4.6 as the base model. Photo: Shutterstock

Zhipu AI sees tenfold surge in overseas users as Chinese AI gains traction

Mobile phone shows the website of Beijing Zhipu Huazhang Technology (Z.ai). Photo: Shutterstock 

Z.ai has seen the number of paid users overseas increase tenfold over the past two months. Photo: Shutterstock

Chinese start-up MiniMax launches record-breaking AI model, challenges Google DeepMind

MiniMax is one of China’s top generative AI start-ups. Photo: Shutterstock

Tech

Tech Trends

Big Tech

Business

News

China

Science

China Future Tech

Companies

China Business

As AI systems grow more powerful, their potential to endanger public safety and social stability has raised alarms


As AI systems grow more powerful, their potential to endanger public safety and social stability has raised alarms.


Chinese AI models comparable to US ones in frontier risks, study finds

.css-1c6uqr6{color:inherit;font-weight:inherit;font-size:inherit;font-family:inherit;line-height:inherit;overflow-wrap:break-word;}As AI systems grow more powerful, their potential to endanger public safety and social stability has raised alarms

As AI systems grow more powerful, their potential to endanger public safety and social stability has raised alarms