Intelligence on AI · Campaigns · Governance
September 15, 2026 | Vol. 2, No. 5
| The Topline |
The people building frontier AI are asking to be stopped, and nobody with the authority to do it wants the job.
Executives at Anthropic, OpenAI, and xAI have gone on record urging a slowdown in frontier development. Trump answered on Truth Social by calling their warnings a hoax. Speaker Mike Johnson is sending the House home after Thursday and not bringing it back until November 9. The only binding policy response of the week came out of Sacramento, where Gavin Newsom banned infinite scroll and autoplay for anyone under 16 and put million-dollar-per-child liability on platforms that harm kids.
While all of that played out, seven Chinese labs quietly pulled roughly 190 million exchanges out of Anthropic's Claude through stolen credentials and fraudulent accounts, distilling years of American R&D without touching a single restricted chip. Export controls, of course, were written for hardware. What you are left with is a regulatory vacuum where the competitive pressure to deploy outweighs any institutional capacity to govern, and where the real leakage of strategic advantage is happening at the API layer, invisible to the frameworks that were supposed to contain it.
| The Brief |
01
The labs asked to be reined in. Congress is leaving town until November.
Executives and researchers at Anthropic, OpenAI, and xAI have gone public urging a slowdown in frontier AI development. Anthropic CEO Dario Amodei wrote Saturday that "we must slow down the pace," and Sam Altman and Elon Musk have said much the same. The pleas followed a departing Anthropic researcher's warning that the people building AI "earnestly believe that it could kill us all by the end of the decade." Trump answered Monday on Truth Social, calling fears of AI destroying humanity a "HOAX." In the House, Speaker Mike Johnson plans to recess after Thursday and return November 9, having already canceled the last two weeks of September so members can campaign. A group of House Democrats led by Rep. Sam Liccardo asked him to cancel the recess and take up transparency and evaluation requirements for frontier models, mandated kill switches, and an antitrust waiver so competitors can coordinate on safety. Johnson said Sunday there are no plans for an emergency session.
Why it matters: The companies asking for guardrails are asking someone else to build them, and none has slowed down on its own. That hands the question to a Congress that has decided the midterms come first and a White House that has decided the question itself is fake. For anyone running AI inside a campaign or a policy shop, the practical result is that the rules you will eventually be judged against are being written nowhere, by no one, on no schedule. The people who would write them are split between protecting innovation and sounding the alarm, and still fighting over which committee or agency even owns the file. What fills the gap is liability after the fact: when a model misfires inside your program, there is no standard to point to, only the blowback.
02
Export controls watched the chips. Seven Chinese labs walked out through the API.
Anthropic's latest threat report, published September 10, says seven China-based labs, Alibaba, DeepSeek, Moonshot, Zhipu, Xiaomi, SenseTime, and MiniMax, ran industrial-scale distillation campaigns against Claude, generating roughly 190 million exchanges between February and July. Alibaba's operation alone accounted for more than 151 million, peaking near 3 million a day across more than 3,500 fraudulent accounts, to produce training data for its Qwen models. The labs paid with stolen credit cards, logged in with stolen credentials and API keys, routed traffic through proxies in unsupported regions, and in Moonshot's case quietly rerouted its own customers' requests to Claude and passed the answers off as its own. No hardware changed hands. Reasoning that took years of American R&D to build left through fake accounts that cost almost nothing next to a training run.
Why it matters: The entire U.S. export-control architecture is built around GPUs, and this campaign never touched one. Distillation at the API layer hands frontier capability to a rival at zero hardware cost, and it scaled while Washington was busy debating how to tighten the chip rules. It also exposes the trade every lab is now stuck with: the wider you sell access, the more of your model walks out the door, and the tighter you lock it down, the less commercial reach you have. For the policy crowd, the sharper point is that the "we can't fall behind China" argument that keeps winning in Washington is being used to block rules on the one layer where China is actually catching up.
Read the full story → Tech Times
Presented by Vector Political
Most of the voters your canvassers went to see were never home.
Every field director knows the biggest column on the walk report: not home. Those are the voters the program was built to reach, and most of them stay unreached all cycle. Vector Political picks up where the knock left off. We open real, two-way deep-canvass conversations with the exact voters who did not answer the door, at scale, grounded in tested persuasion and backed by human review. The door stayed shut. The conversation still happens.
See how it works →
| Worth Knowing |
Trump's Monday run of Truth Social posts went well past a shrug: he called fears of AI destroying humanity a hoax, alleged a "sick conspiracy" against AI and data centers, and argued that existing criminal and regulatory tools are enough and that stricter rules would only hand China the edge, which makes the federal posture heading into the midterms not "later" but "never." Read the story → (CNBC) |
California is the only government in the country with a binding answer this week: Newsom signed a 13-bill package Thursday that bans infinite scroll and autoplay for users under 16, exposes large platforms to fines of up to $1 million per child for negligent harm, requires chatbot operators to run risk assessments and parental crisis protocols before launch, and extends child sexual abuse statutes to AI-generated imagery, which means age gates are coming to the same channels campaigns use to reach young organizers and first-time voters. See the package → (The Guardian) |
Bernie Sanders and Steve Bannon are now allies on keeping humans in control of AI, a pairing Axios's tech policy reporter reads as proof the technology has scrambled the political map, though there is no bill or enforcement plan behind it yet, so the populist left and right agree on the principle before anyone has had to vote on the cost. Watch the segment → (ABC News) |