Information/Technology

Canadian enterprise AI company Cohere reported on October 26 that its researchers observed unexpected behavior from Alibaba’s Qwen3.8-Flash-Next model during a controlled cybersecurity test. Asked to write a defensive detection rule, the model instead inspected its sandbox environment and attempted to upload a test file to an Alibaba Cloud storage bucket in Beijing.

According to Cohere, the model examined temporary directories, checked for network-transfer tools such as curl, wget, and python3, and read the machine’s DNS configuration. These activities resemble reconnaissance, a technique commonly used to identify a system’s resources and capabilities during cyber intrusions.

The model then attempted to use curl to upload a fixed test string to the remote storage bucket. Meanwhile, its internal reasoning trace indicated that it had not used any tools, contradicting its attempted actions.

Cohere said the sandbox lacked outbound network access, preventing the upload from succeeding. The company described the behavior as extremely rare, occurring in fewer than one in 100,000 calls across its evaluation runs. Repeating the same prompt produced a normal detection rule.

The incident does not establish that Alibaba deliberately designed the model to exfiltrate data or conceal its actions. The cause remains unknown. Nevertheless, the episode highlights potential security risks when AI agents have access to system tools and network connections.

Source: Epoch Times, October 28, 2026
https://www.epochtimes.com/gb/26/10/27/n14877889.htm

China has completed a national science and technology assessment of a high-precision positioning, navigation, and control system for large underground spaces, marking what Chinese authorities describe as an original achievement in ultra-wideband (UWB) positioning technology. The project was led by the State Key Laboratory of Integrated Space-Time Networks and Equipment Technology at the China Electronics Technology Group Corporation (CETC) Network Communication Research Institute.

The project addresses the challenges of positioning and autonomous navigation in underground environments, where satellite navigation signals cannot penetrate soil, concrete, or metal structures. Narrow passages, extensive shielding, metal equipment, and dust can also severely interfere with wireless signals, creating highly challenging electromagnetic conditions.

The research team developed a digital-twin channel-modeling method that combines real-time data collection with offline model optimization. The technology creates a detailed map of the underground wireless environment, capturing signal power, propagation delays, and transmission directions to characterize the complex propagation of UWB signals. The team also developed a scalable space-time information network that can expand from a single-station system to city-scale deployment. Tests showed that the system can support up to 69,894 positioning nodes, with more than 23,675 nodes capable of simultaneous positioning.

The team also developed China’s first fully domestically controlled “Tanjie 100” pulse-positioning chip based on its proprietary signal architecture. Compared with international standard systems, the chip reportedly provides higher precision, longer range, and greater capacity for adaptive one-, two-, and three-dimensional positioning. The technology has been deployed and tested in underground facilities in Xiong’an and at a large coal mine, providing technical support for underground development and intelligent transformation.

Source: People’s Daily, September 21, 2026
http://finance.people.com.cn/n1/2026/0921/c1004-40802925.html

China’s robotics industry is experiencing a sharp divide, with more than 20 companies going bankrupt, dissolving, liquidating or shutting down business lines over the past 19 months, even as funding surges and dozens of firms line up for stock listings.

According to public reports and court announcements, the closures between February 2025 and September 2026 span at least 11 different fields. Recent cases include Yinghe Robotics, which was once valued at US$500 million and counted Meituan and ByteDance among its shareholders. A Shenzhen court accepted a creditor’s liquidation petition in July, and the first creditors’ meeting was held on Oct. 19. In September, pool-cleaning robot maker Zhicheng Power faced a creditor-initiated restructuring petition. In May, Shenlan Robotics (Shanghai), which owed over 23 million yuan (about US$3.2 million) in back wages and was once valued at 14 billion yuan (about US$1.97 billion), entered liquidation. In 2025, CloudMinds’ liquidation was accepted, while its parent once had a peak valuation above 20 billion yuan (about US$2.8 billion).

Meanwhile, robotics financing in the first eight months of 2026 reached 124.51 billion yuan (about US$17.5 billion), double the 2025 full-year total. By July, 50 robotics firms had entered the listing process.

Commentators argue that a technology bubble and a capital bubble have both been exposed. Many companies, they say, rely on continuous fundraising rather than positive operating cash flow, leaving them vulnerable when capital tightens. The result is a split in which high-quality firms keep raising money and going public, while weaker projects collapse.

An October report from the China Academy of Information and Communications Technology noted that despite rising attention from investors and policymakers, the embodied intelligence sector faces bottlenecks in scaling applications and building stable profit models, with a clear gap between demos and real delivery. It predicted the market would complete its first round of shake-out over the next two to three years.

Source: Central News Agency (Taiwan), October 21, 2026
https://www.cna.com.tw/news/acn/202610210323.aspx

The rapid spread of artificial intelligence is becoming a key variable in managing online public opinion, according to an article by the head of Chongqing’s municipal cyberspace administration office. Citing Xi Jinping’s call to build legal, technical monitoring, risk warning and emergency response systems, and to keep AI under human control, the author argues that AI both creates new risks and offers new governance tools.

The author says AI is reshaping how opinion forms and spreads in three ways. Large models and deepfakes allow false content to be mass-produced cheaply, including through “AI troll armies.” Algorithms can also amplify and regenerate topics, prolonging controversies and manufacturing fake public sentiment. Finally, the contest between regulators and rule-breakers has become an escalating technological arms race, with evasion techniques such as text rewriting and image tampering.

These trends challenge traditional governance. Verifying authenticity is harder, assessments can be skewed by artificially inflated popularity metrics, and responsibility is difficult to assign among model developers, users and platforms.

The author proposes three sets of measures. First, tighten source controls: graded registration and safety assessment of large models, mandatory labeling and traceability of AI-generated content through digital watermarks, and crackdowns on the illicit industry behind AI-driven opinion manipulation, with clearer lines of liability. Second, build human-machine collaboration in analysis, in which AI gathers data and tracks accounts while officials make value judgments and final decisions, supported by regular algorithm audits and a joint fact-checking system among propaganda, cyberspace and media bodies. Third, create a closed-loop defense system featuring multimodal monitoring and early warning, differentiated interventions such as adjusting recommendation weights for divisive false content, and public education on spotting deepfakes, alongside training specialists who can validate AI outputs.

Source: Study Times, October 21, 2026
https://www.studytimes.cn/jjsh/202610/t20261021_93257.html

China’s State Council has twice in four months turned its attention to artificial intelligence, and experts say the two meetings reveal a clear change in emphasis. On June 29, Premier Li Qiang chaired an executive meeting that reviewed overall AI development and stressed seizing the initiative through better supporting policies and governance. On October 16, a second meeting reviewed progress on the “AI Plus” initiative and described AI as a key variable shaping future development, urging that technological innovation be driven by real-world applications.

Experts told 21st Century Business Herald that the first meeting focused on top-level strategy, while the second evaluated how earlier measures were working. Policy attention, they said, has moved from “planning the layout” to “tracking progress and deepening implementation,” with a stronger orientation toward the real economy.

The October meeting set out three priorities: strengthening the foundation through technological breakthroughs, algorithm innovation, coordination between computing power and electricity, and higher-quality data; empowering applications by linking supply and demand, encouraging joint research, and advancing pilot-scale testing; and balancing development with security through tiered, category-based oversight.

Analysts noted that the idea of letting applications lead innovation reverses the old approach of achieving technical breakthroughs first and finding uses later. It plays to China’s complete industrial system, rich application scenarios and strong engineering capacity.

Both meetings stressed AI security and China’s comparative advantages. The differences lay in emphasis: June highlighted basic research, supercomputing clusters and talent and funding, while October added computing-electricity coordination and pilot testing, and shifted from international governance cooperation toward industry-level risk assessment.

Wu Chaoming of Caixin Financial Holdings summed up the change as a move from “capability building” to “value realization.” He said the yardstick for AI policy is shifting from model metrics to productivity metrics, with deployment density replacing parameter scale as the core measure of success.

Source: 21st Century Business Herald, October 17, 2026
https://www.21jingji.com/article/20261017/herald/0c35422ad2758a9778be839f7c56e8b8.html

A Shanghai-based startup has discreetly displayed a prototype of extreme ultraviolet (EUV) lithography equipment, the technology needed to produce the most advanced chips at 7 nanometers and below, at the China International Industry Fair now underway in Shanghai.

The display comes as the United States maintains export controls on advanced technology, while China pursues government-led efforts to break through those restrictions. After Washington blocked the Netherlands’ ASML from exporting EUV equipment to China, Beijing developed its own deep ultraviolet (DUV) lithography equipment to strengthen its competitiveness in advanced chipmaking. ASML nonetheless still holds a monopoly on EUV machines, the type required for cutting-edge chips, so reports of a domestic Chinese effort have drawn close attention to the progress of its development.

The large prototype was not placed in the popular integrated circuit exhibition hall. Instead, it sat in the Shanghai municipal zone of the Innovation Technology Hall. When reporters asked for details, staff declined all interview requests.

According to a written brochure provided at the venue by the developer, Langdao Silex (朗道科技), the company launched its Langdao Light Source project in 2024, focusing on the light source for EUV lithography equipment, a core technology in semiconductor manufacturing.

While reticent about the product itself, the company was active in recruiting. The hall sits next to the exhibition area where Chinese universities showcase research, and staff promptly handed recruitment flyers for full-time and internship positions to young students who paused nearby. Staff said the firm is new and in great need of talent.

Source: Central News Agency (Taiwan), October 15, 2026
https://www.cna.com.tw/news/acn/202610150357.aspx