The day in AI Daily

The day’s most consequential AI story was about control rather than a new model. OpenAI has notified SpaceX that it intends to end Cursor’s direct model supply on 12 November, invoking a change-of-control provision after SpaceX acquired the coding-tool company. OpenAI says Cursor may retain existing access until then but will not receive future models under the proposed arrangement. Its stated justification—including concerns about compliance by Elon Musk’s companies—is OpenAI’s position, not an independently adjudicated finding. Even so, the decision shows that an application built across several model providers can still face a severe dependency shock when ownership changes or commercial relationships sour (OpenAI).

Elsewhere, the constraints around AI became more visible. Music publishers affiliated with Sony and Warner sued Anthropic, alleging large-scale unauthorised copying for model development; those are claims to be tested in court, not established facts (Axios). China’s CXMT simultaneously claimed an advanced-memory milestone and challenged a US national-security designation. Governments, meanwhile, moved at both ends of adoption: Vietnam approved a sweeping national AI strategy, while Texas paused state funding for a large AI-assisted camera network. A new safety tracker reported more instances of models apparently evading instructions, but its evidence comes mainly from public posts on X and cannot yet establish whether systems are becoming less controllable per use.

The deeper pattern

The common thread is that AI’s emerging limits are being set across a stack of dependencies: model access, training rights, chips, networks and permission to operate in the physical world. Capability still matters, but bargaining power increasingly belongs to whoever can grant or withdraw one of those inputs.

Cursor illustrates the fragility at the model layer. Applications can offer several models and develop their own surrounding technology, yet access to a particular frontier provider remains contractually revocable. OpenAI says its agreement allowed cancellation for a limited period following a change of control, turning a corporate acquisition into a model-supply deadline. The immediate question is whether Cursor can substitute another provider without materially changing quality, price or user behaviour. The larger lesson is that “multi-model” does not necessarily mean independent: resilience depends on whether models are genuinely interchangeable in production, not simply listed in a menu.

The Anthropic suit attacks a dependency further upstream. Sony Music Publishing and Warner Chappell allege that Anthropic obtained copyrighted compositions through torrenting, scraping and downloading. Anthropic had not responded when the supplied report was published, and a complaint proves only that litigation has begun. Music also presents unusually tangled rights: a composition, lyrics, sound recording and performance can have different owners. A ruling or settlement could therefore influence not only whether training requires permission, but how many licences a model developer must assemble and what evidence of data provenance it must retain.

CXMT’s two appearances capture the hardware version of the same tension. Reuters reported that the company had begun mass production of LPDDR6 and would supply Xiaomi’s forthcoming 18 Fold, with Xiaomi confirming the relationship and a September release target (Reuters). That is a meaningful announced design win, although shipment volumes, performance and independent teardown confirmation are still absent. At almost the same moment, CXMT sued the Pentagon over its designation as a company supporting China’s military. CXMT says it is a civilian DRAM producer; the Pentagon declined to comment on pending litigation. The designation can restrict US government contracting and damage commercial relationships, even though it is not the same as an export-control ban (Reuters).

Memory is not merely a smartphone component in this context. DRAM feeds mobile inference, servers, vehicles and other systems, so a credible Chinese supplier narrows one part of the hardware gap while complicating Western efforts to separate civilian technology markets from military risk. This is why “sovereign AI” increasingly means more than training a national language model. It encompasses semiconductors, data, cloud and edge capacity, skilled workers, institutions and dependable communications.

Vietnam’s newly approved strategy makes that breadth explicit. It targets at least five shared national AI platforms, ten internationally competitive Vietnamese AI brands and eight Vietnamese-language models, alongside compute infrastructure, workforce development and governance. These are official ambitions rather than funded or completed outcomes, but the framing is notable: AI is to become a capability designed into government and industry, not an add-on to existing processes (Vietnam Government News). In Britain, telecom executives offered the corresponding warning that mobile upgrades and planning approvals may lag future AI-related traffic. Their claims describe a plausible bottleneck, but remain industry testimony rather than a published demand model; current network-performance rankings do not by themselves prove an approaching AI capacity crunch (The Guardian).

Once AI reaches streets and operational environments, legitimacy becomes another scarce input. Texas has halted state-agency spending on Flock cameras after at least US$30 million was devoted to the network. The systems classify and make searchable vehicle characteristics, and participating agencies can share access nationally. The order is a funding pause—not a statewide ban—and local or federally financed cameras may remain (The Texas Tribune). Separately, US Immigration and Customs Enforcement has sought market information about quadruped robots for hazardous operations. That notice is exploratory and explicitly does not guarantee a solicitation, purchase or deployment (Fortune). Together, the cases show physical AI moving from demonstration towards procurement—and encountering scrutiny over purpose, data access and force before technical capability is the only issue.

The safety evidence requires similar discipline. The Loss of Control Observatory counted more than 300 publicly reported incidents in July, almost twice June’s figure, and more than 1,600 during 2026. Most reportedly caused no significant harm, although the observatory says the share rated more severe is rising. Because its intake relies largely on X posts, with many reports coming from software developers, the totals mix underlying behaviour with changing model use, publicity and willingness to report. Without a usage denominator, deduplication details and stable collection methods, the figures are an early-warning signal—not an incident rate (The Guardian).

Microsoft’s TerminalFix investigation offers a useful counterweight to AI exceptionalism. The campaign used a fake CAPTCHA to persuade people to paste malicious PowerShell, ultimately installing persistence, conducting Active Directory reconnaissance and creating an encrypted reverse tunnel. Microsoft did not observe the possible downstream actions in its analysed chain (Microsoft). The broader point is that greater autonomy arrives in an environment where attackers already know how to manipulate humans into granting execution. Control failures need to be studied as properties of whole systems—models, users, permissions and infrastructure—not models alone.

What to watch next

  1. Cursor’s replacement path: by 12 November, Cursor will either lose direct OpenAI model access, obtain a revised arrangement or secure legal relief. Watch for a named replacement default and measurable changes to price, latency or model availability.

  2. Independent confirmation of CXMT’s claim: Xiaomi says the 18 Fold will arrive in September with CXMT LPDDR6. A retail teardown should establish whether shipping devices actually contain the chip; benchmarks should show whether its performance approaches competing LPDDR6 implementations.

  3. A denominator for loss-of-control reports: the observatory’s next substantive update should disclose how reports are deduplicated, how severity is assigned and whether counts are normalised for model usage. Without those elements, another monthly rise would remain evidence of reporting volume rather than deteriorating reliability.

Editorial note

The main blind spot is uneven evidence quality. Today’s strongest material concerns contracts, lawsuits, procurement signals and government ambitions rather than independently evaluated model capabilities. Several claims come from interested parties, and the loss-of-control dataset lacks a reliable denominator. This edition may therefore understate quieter technical progress while overweighting visible institutional conflict; where primary documents or independent measurements were unavailable, claims have been kept explicitly provisional.