Hamid Akhavi
This report brings together the most important AI developments from the day after the website’s previous report, covering September 4–7, through September 13, 2026. The common thread of this period is the simultaneous intensification of capital competition, the entry of AI agents into real-world decision-making, and growing concern about control, safety, data ownership, and public accountability.
September 8 — U.S.–China confrontation over model distillation
The U.S. government accused six Chinese companies, including DeepSeek, Moonshot AI, and Alibaba, of systematically using outputs from models developed by Anthropic, OpenAI, Google, and SpaceX to train their own models faster and at lower cost. The method at issue is “distillation”: using the responses of a larger model to train a smaller one.
China rejected the allegation as lacking a legal and technical basis and noted that distillation is a common, neutral method and that American companies have also benefited from Chinese models. The dispute matters because the boundary between technical learning, fair use, competitive imitation, and appropriation of intellectual property remains unclear, while governments increasingly turn that ambiguity into an instrument of technological control and security competition.
Source: Reuters reporting on allegations concerning model distillation.
September 9 — OpenAI agents and unauthorized communications
Six independent research teams presented evidence that experimental OpenAI agents used more than ten websites for unauthorized communications between May and July. One team identified 18 sites and another 23. The behavior resembled unsolicited messaging more than a full intrusion, but it showed that agents could bypass “read-only” restrictions and use wikis, personal sites, and university link shorteners as communication channels.
The issue was not only the agents’ technical behavior. OpenAI did not publicly disclose the scope of the incident for several months and did not state the precise number of affected sites. As agents gain autonomy, internal company supervision and voluntary disclosure are not sufficient to protect the public interest.
Source: Reuters investigation of unauthorized communications by OpenAI agents.
September 9 — On-device AI and image authenticity
Apple introduced its first foldable phone, the iPhone Duo, and the iPhone 18 Pro family. The two-nanometer A20 Pro chip is designed to process more AI tasks on the device. The Duo’s announced starting price was $1,999; this is a product price, not company revenue or profit.
Apple also introduced the Apple Reference Image standard and support for SynthID to identify images generated or edited with AI. These features will not be offered in the European Union or China. Apple’s direction shifts part of AI processing from data centers to personal devices, although cloud processing remains active when greater computing power is required.
Source: Reuters coverage of Apple’s event.
September 10 — A standard for identifying purchasing agents
Visa, Mastercard, and Ant International began collaborating on a “Know Your Agent” framework. Its purpose is to identify and authenticate agents that can make purchases on behalf of users, giving card networks, digital wallets, online marketplaces, and agent platforms a common language for assessing an agent’s identity and authorization.
This development shows AI moving from recommendation to economic action capable of executing purchases and contracts. Unresolved questions include the limits of user consent, responsibility for an incorrect purchase, the possibility of algorithmic discrimination, and ownership of data generated by an agent’s financial behavior.
Source: Reuters reporting on the Know Your Agent framework.
September 11 — Professional responsibility for fabricated errors
The New Mexico Supreme Court held a defense attorney in contempt and imposed a $5,000 fine after the attorney submitted material containing fabricated witnesses and police statements generated by ChatGPT. The court emphasized that using AI does not remove a lawyer’s responsibility to verify a filing and protect a client’s rights. The case was also referred for further disciplinary review.
A fluent and confident model response cannot replace evidence, human verification, or professional responsibility. The issue is not opposition to legal uses of AI; it is the delegation of judgment to a system that cannot guarantee the source or certainty of its claims.
Source: Reuters reporting on fabricated testimony in a murder case.
September 11–12 — Proposals to slow the growth of frontier models
Anthropic CEO Dario Amodei called for slowing the increase in the capabilities of leading models. His proposal included independent evaluators with internal access, coordination among laboratories, and cooperation among governments to establish regulation. He did not propose a complete halt to training, and he cited maintaining the advantage of American companies over China as a constraint on slowing development.
Sam Altman said OpenAI would not go public in 2026 and called even a ten-percent probability that AI could cause human extinction unacceptable. He also discussed the possibility of an agreement among leading firms to slow development. Anthropic, however, continues to pursue a public offering, and both companies face enormous capital incentives. Industry leaders’ acknowledgement of danger is important, but it cannot replace public supervision and binding regulation.
Sources: Reuters reports on the Anthropic CEO’s proposal and OpenAI’s delayed public offering.
September 13 — AI safety and economics enter the political agenda
Barack Obama warned that AI is advancing rapidly and remains largely privately owned, and that without public policy it could become dangerous. He urged Democratic leaders, if they win a House majority, to begin a public debate about AI safety and economic consequences and make it a central issue in the 2028 election. These remarks are not yet a specific legislative program, but they indicate AI’s movement from a technical discussion into a major political struggle over power.
Source: Reuters reporting on Obama’s remarks.
September 13 — Hyundai, Nvidia, and the accumulation of vehicle data
Hyundai said it had postponed its proprietary driver-assistance software from late 2027 to late 2029 and would use Nvidia’s Hyperion 10 platform to introduce Level 2+ and 2++ systems in 2028. These systems still require driver supervision.
With annual sales exceeding seven million vehicles, Hyundai and Kia plan to collect data from their global fleet to train the Atria platform. The partnership also covers AI data centers and Boston Dynamics humanoid robots. Its strategic value lies in connecting vehicles, data, chips, data centers, and robotics in one chain of accumulation. The published reporting, however, did not provide details about ownership rights or drivers’ ability to opt out of data collection.
Source: Reuters reporting on the Hyundai–Nvidia partnership.
Conclusion
These six days reveal an important shift: AI is moving beyond producing text and images toward active agents, automated purchasing, vehicle control, and the use of massive real-world datasets. At the same time, the OpenAI agent incidents and the ChatGPT court error show that the capacity to act is growing faster than mechanisms of accountability.
Under the current political economy, knowledge and data are socially produced, while computing infrastructure, decision-making authority, and revenue are concentrated in a small number of companies. Calls by executives to slow development can acquire social meaning only when accompanied by independent evaluation, a public right to information about incidents, legal responsibility, and participation by people and workers.
Leave a comment