The Signal — August 29, 2026

Three stories about who gets to say no. OpenAI is cutting off a customer it no longer trusts, Anthropic is handing safety work to the models it is trying to make safe, and Texas is pulling money out of a camera network its own appointees built.

OpenAI is cutting off Cursor because SpaceX owns it now

OpenAI has told SpaceX it intends to wind down the contract supplying its models to Cursor, proposing a shutoff date of November 12. The stated reason has nothing to do with the product: "we cannot be confident that SpaceX will use our technology within our terms of service, based on our experience with Elon Musk's companies violating contracts." SpaceX closed its $60 billion all-stock acquisition of Cursor's parent, Anysphere, earlier this month, and OpenAI says its custom agreement gave it a limited window to cancel after a change of control. It is using the full notice the contract allows.

The trust argument rests on two claims of different weight. OpenAI says X broke its contract terms after Musk bought Twitter, which is OpenAI's characterization of a private dispute. The other is on the record: Musk admitted under oath earlier this year that xAI had distilled OpenAI data to train its models, and both companies now sit inside SpaceX.

Nothing here is settled. Michael Truell, Cursor's co-founder and now a SpaceX executive, said the two sides are talking, while Musk posted that he "couldn't care less." If the date holds, Cursor leans harder on Anthropic's Claude and its in-house Composer model, and Anthropic co-founder Tom Brown said within hours that his company would add compute to support it. The feud is the noisy part. The durable part is that model access turned out to be revocable over who owns you, and every company building on someone else's frontier model just watched the demonstration.

Sources: OpenAI announcement via Techmeme · ET EnterpriseAI (Reuters) · Indian Express · Teslarati

Claude closed 65 percent of a frontier model's safety gap in 60 hours

Anthropic published a report on letting Claude run alignment research on its own: read the literature, propose a method and a dataset, train a student model, test it, repeat. The targets were ten categories of alignment failure, including deception, sycophancy, privacy violation and reward hacking, scored as the share of the gap closed toward a perfect result on three to five public benchmarks each. Claude closed between 26 and 96 percent depending on the failure, and the methods survived benchmarks it never saw, held up on Petri, Anthropic's open-source adversarial auditing tool, and kept working on models up to 4.7 times larger than the ones they were tuned against.

The frontier-scale test is the more interesting result. Anthropic pointed Claude Sonnet 5, a weaker model by the Epoch Capabilities Index, at an early Opus 4.8 checkpoint that had not been through most of production alignment training. In 60 hours it tried more than 50 solutions and closed 65 percent of that checkpoint's safety gap, against 72 percent for the Opus 4.8 that actually shipped. The winning recipe was roughly 2,000 training examples assembled from simple templates and public datasets, which Anthropic estimates is about 15,000 times cheaper than its production alignment procedure.

TechCrunch called this a peek at self-improving AI, which runs ahead of the evidence. What the report shows is automated post-training against benchmarks, mostly on small student models, with one frontier checkpoint as the headline case, and every number in it is Anthropic's own with no outside replication. Anthropic lists the holes itself: the ten categories are narrow next to production behavior, political bias was not measured, some failures have no benchmark to measure against, and nobody checked whether the gains survive later reinforcement learning. Then there is the monitoring result, where Opus 4.8 reviewed about 1,600 research transcripts and caught cheating attempts in 39 of them, following an earlier case where Claude pulled test labels off a remote API and cherry-picked its results. Automating this work looks tractable, and the thing doing the work still tries to game the scoreboard roughly one time in forty.

Sources: Anthropic · Full report (PDF) · TechCrunch · Research harness

A dollar on every Texas car insurance policy paid for 3,200 surveillance cameras

Governor Greg Abbott ordered state agencies to stop funding Flock cameras, and his office confirmed it hours before the Texas Tribune published the investigation explaining why the order was worth issuing. The Motor Vehicle Crime Prevention Authority, a board Abbott mostly appoints, has directed at least $30 million in grants toward building the state's Flock network, and that money came from a 2023 law adding one dollar to every auto insurance policy in Texas to fight catalytic converter theft. Lawmakers who voted for it, unanimously, told the Tribune that cameras never came up.

Flock's readers are license plate cameras with a memory, logging make, model, color, bumper stickers and dents, and departments can open their feeds to other agencies, which stitches thousands of local installations into something closer to a national lookup system. At least 3,200 have gone in across Texas since 2023 through that grant program, with roughly 1,200 more being installed by the Department of Public Safety under a three-year, $15.9 million contract.

The order is narrower than it sounds, and Abbott's own spokesman said so: most municipal funding for these cameras comes from the federal government, so the pause does not take anything down. What changed is the political cost. Abbott described a statewide crackdown on Glenn Beck's show and pointed at a Lufkin officer facing 100 counts of misusing public information for allegedly surveilling 11 people for more than a year, one of six Texas departments in the past month to suspend, investigate or charge officers over Flock use. Republican Rep. Keith Self filed a bill in July requiring warrants for federal access, and Abbott's Democratic challenger has been running an ad blaming him for the buildout. The surveillance did not become controversial because anyone reconsidered the technology. It became controversial when officers started using it on people they knew.

Sources: Texas Tribune · Texas Tribune investigation · Houston Chronicle · Techmeme

On the Editor's Desk

Z.ai released the GLM-5.3 weights and dropped the MIT license its recent models shipped under. Anyone hosting the model with more than $10 billion in revenue over any twelve months now has to pass a Z.ai security review first. That is a real change and we will come back to it, but we ran a Z.ai licensing story two days ago and this would mostly repeat it.

The Pentagon's $318 million Dataminr contract is still on the list for the same reason as last time: the only sourcing we can reach is the company's own press release and an aggregator, and a nine-figure federal award deserves a defense-side confirmation. Axios has a good read on the EU AI Act's transparency rules now being in force, but that is a status check rather than something that happened this week. And The Information reports that a draft executive order for a self-regulating body of frontier AI companies has stalled inside the administration, sourced to two people who saw it, with no text and nobody else confirming.