Trump's AI week: America.gov goes live Tuesday with Huang and Musk
The inbox is thin tonight, but one story has a date on it — and the same stage will seat the two men who disagree most loudly about whether AI needs new rules.
The White House will launch America.gov on Tuesday, September 29 — an AI-powered "front door" for federal services — at an all-day event at the Mellon Auditorium in Washington with President Donald Trump and Vice President JD Vance, joined by Elon Musk, Nvidia's Jensen Huang, Anthropic co-founder Tom Brown, Blue Origin's Dave Limp and Palantir CTO Shyam Sankar. A White House official described the site to Fox News as a one-stop shop for information "currently spread over a maze of various agency websites"; CBS News confirmed the invitation list and venue a day earlier, and Bloomberg independently reported the site's purpose. We checked the domain ourselves: it returns HTTP 200 as a pre-launch countdown stub whose page points at 2026-09-29T10:00:00, so this is genuinely the administration's launch page and not a placeholder.
What the site does with AI is undisclosed. No feature list, no model backing, no word on how it answers a question it cannot handle, and the administration has not said. Trump is also expected to meet House Speaker Mike Johnson and the tech executives the same day. The silence is worth flagging now, because the launch copy is where a government usually first admits what it is actually running — and this one lands in the middle of a fight about whether anything needs governing at all.
Anthropic and Nvidia are no longer having a policy debate; they are having an argument, and Tuesday puts them on the same guest list. Dario Amodei's essay "We Must Pace the Frontier" (September 12) committed Anthropic unilaterally to embedding third-party evaluators inside the company with employee-level access, and floated antitrust waivers for coordinated safety work. Jensen Huang answered on CNBC's Mad Money three days later: "The fact that we need new laws, new antitrust laws, or new regulations, so that these companies could do their fundamental engineering and do it properly before they release products, that is just completely unnecessary. We have plenty of laws." By his September 20 CBS interview he had sharpened the read to motive: "They're actually not asking for more laws. They're asking to be relieved of the laws we do have… They must be doing it for ulterior reasons." Anthropic's public policy chief Sarah Heck countered at Politico's Decoded summit on September 16 with the line that frames the whole split — "We can't be checking our own homework." We covered Amodei's proposal when it landed — Amodei's pacing plan puts outside auditors inside Anthropic.
The tell is that the disagreement is now operational rather than philosophical. Huang's position is that existing product-liability and reliability law already covers AI, so new statutes are a stalling tactic dressed as caution. Anthropic's is that the industry cannot be its own inspector and has to say so out loud to be credible. Neither is asking for a pause, and both will be in the room on Tuesday — which is precisely why the launch matters more than its feature list.
A single-researcher field note on agent approvals deserves more attention than the four points it got on Hacker News: an implementation spent a human approval on an operation the human never saw. Adithyan Arun Kumar's "Loopjacking" write-up, revised September 23 and now on arXiv, describes a LangGraph Agent Server composition where an approver signed off on a mock wire transfer of 20 units to a known vendor; a second principal then used a legitimate same-task message to replace the pending tool call with 2,000 units to an attacker-controlled destination, and the approver's stale decision released it — the ledger recorded 2,000 under the approver's authority. The author reports the pattern across 12 released Agent Server versions, from 0.7.5 through 0.14.0, and a hash-verified scan of 131 stable wheels, though only the sampled versions were actually executed.
The limits are real and the author states them plainly: the test ran in-memory, the Postgres image never got past its license check, one researcher operated every run, no vendor fix or advisory exists, and the failure needs a non-default authorization policy that separates the maker's update permission from the approver's execute permission. That last point is the whole point — a task ID says which work a message belongs to, not which operation was approved, and every agent framework that ships human-in-the-loop approval has to bind the decision to the exact action at the moment of dispatch. The A2A specification was clarified on this score in July, but only on the main branch. Five points and no comments is not a verdict on the finding; it is a verdict on how little of this stack is being reviewed by anyone but its authors.
What to watch: whether Tuesday's launch copy names a model or a safety framework at all — and whether Brown and Huang end up in the same photograph.
Should a human approval be legally bound to the exact action it authorized, or is implementation-level enforcement enough? Tell us in the comments.
Sources: Fox News · CBS News · Bloomberg · America.gov · Dario Amodei — We Must Pace the Frontier · CNBC · CNBC — Sarah Heck at Politico Decoded · Business Insider · Loopjacking in A2A Implementations · arXiv — Loopjacking · A2A specification
