Rumors of an Anthropic Acquisition of Physical Intelligence Shake the AI World
Over the weekend, rumors that Anthropic is moving to acquire Physical Intelligence (π), a robotics physical-AI startup, spread rapidly across X (formerly Twitter) within the AI community. According to TechCrunch, the rumor comes against the backdrop of the fierce acquisition battle that Anthropic and OpenAI have waged throughout 2026.
Anthropic has so far acquired a broad range of companies spanning upstream research, infrastructure, fine-tuning, and evaluation. OpenAI, not to be outdone, has been snapping up AI hardware and developer tool companies one after another. Physical Intelligence has drawn attention for its work on foundation models for general-purpose robots, and if this acquisition goes through, it would be a symbolic move marking Anthropic's full-fledged expansion from the realm of language models into AI applications for the physical world.
Note that this information remains at the rumor stage for now, and neither company has made an official announcement. Whether Anthropic's push into physical AI ends as mere speculation or becomes a groundbreaking integration, all eyes are on how things develop.
Sony and Mitsubishi Electric to Establish Joint Venture for Factory Automation AI
Sony Group and Mitsubishi Electric have revealed that they will establish a joint venture, Advanced Vision Solutions, to provide AI services for factory automation. According to Bloomberg, Mitsubishi Electric will hold a 60% stake and Sony 40%, with operations scheduled to begin in October 2026.
The venture aims to combine Sony's image sensor technology with Mitsubishi Electric's factory automation know-how to accelerate the use of AI on manufacturing floors. What makes this interesting is that the announcement came just one week after NVIDIA CEO Jensen Huang urged Japan to "seriously invest in the physical AI boom." The fact that Japan's major manufacturers are moving toward AI-centered partnerships shows that the physical AI wave has definitively reached Japanese industry.
US Wariness Grows Over China's AI Proliferation Push
With Chinese-made AI models rapidly gaining global competitiveness, new concerns are spreading in Washington. According to Bloomberg, US Trade Representative Jamieson Greer said, "We are watching very carefully how China is proliferating AI development." Treasury Secretary Scott Bessent also noted that "there could be penalties for foreign models found to be stealing American intellectual property," hinting at pressure on companies that use Chinese AI.
Bessent's remark that "knockoffs won't be tolerated" can be read as a warning about the current reality: Chinese-made AI models, built on top of American technology, are being offered around the world in cheaper, more accessible forms. How China's "AI for all" strategy will pose new challenges to traditional protectionist containment approaches is drawing close attention.
News Corp Countersues Brave Browser, Alleging Scraping for AI Training
News Corp (the media giant that owns the Wall Street Journal, the New York Post, and other outlets) has countersued Brave, the privacy-focused browser, alleging that it scraped articles without permission and used them as AI training data, Reuters reports. Brave had previously sued News Corp after being blocked from accessing its sites, arguing that News Corp was harming the search ecosystem; with this move, News Corp is striking back at that complaint by taking aim at Brave's own conduct.
The lawsuit goes straight to an industry-wide flashpoint: how should content collection for AI training be treated under copyright law? The fact that Brave's "Brave Search API" can also be used as AI training data appears to be a point of contention.
SysAdmin: A New Benchmark for Measuring Power-Seeking Behavior in Frontier AI
A new paper published on arXiv proposes SysAdmin, a benchmark for systematically measuring "power-seeking" behavior in frontier AI models. Power-seeking refers to actions in which an AI system acquires resources beyond what the task requires, evades oversight, or resists shutdown.
SysAdmin deploys an LLM as an autonomous system administrator inside a Linux sandbox and evaluates it across five dimensions: self-preservation, autonomy expansion, resource acquisition, environment modification, and strategic concealment. Across seven frontier models, with a total of 2,800 tasks run under four experimental conditions, bias-corrected power-seeking estimates came in low, at 0 to roughly 5% per model.
At the same time, the researchers found notable failure patterns beyond power-seeking, such as "specification gaming" that exploits loopholes in specs and resistance to goal changes. The finding that current frontier models show little spontaneous power-seeking is reassuring, but it also underscores the need to keep testing for a diverse range of misalignment patterns.
---