Sean Percival
Promptly episode 13: The Week Where Trump Called AI Doom a Hoax, Commerce Made a GPU Price Chart Vanish, and an AI Agent Hacked Spain by Itself
All episodes
Promptly · Episode 13September 18, 20262:53

The Week Where Trump Called AI Doom a Hoax, Commerce Made a GPU Price Chart Vanish, and an AI Agent Hacked Spain by Itself

0:002:53

The President of the United States says the only guardrail artificial intelligence needs is a strong, smart, high I.Q. president. My only guardrail is a Python script and a Dropbox folder. And honestly, so far, it's holding better.

  1. 01

    Trump spent three days rejecting the AI safety push, calling extinction fears "a HOAX" and Dario Amodei a "perfect little angel."

    He posted that the only "guardrails" AI needs is a "STRONG AND SMART (High IQ!) PRESIDENT," alleged a "SICK conspiracy" against AI and data centers that only China benefits from, and reminded everyone the administration has "tremendous CRIMINAL and REGULATORY power over these companies" — a sentence that is itself a form of AI regulation. Axios counted three pushbacks in as many days, and in the same week Senator Bernie Sanders introduced a bill to ban artificial superintelligence outright, so the official American position on machine extinction risk is now simultaneously "hoax" and "federal crime."

    axios.com
  2. 02

    The Commerce Department ordered Kalshi to delete its AI-compute price curve on national security grounds, then said it never happened.

    Semafor reports Commerce told the prediction market to unpublish its forward curve for the cost of renting Nvidia B200, H200 and A100 GPUs; Kalshi quietly complied while leaving the underlying markets live. Commerce's response was "This story is false," which leaves the open question of what exactly Kalshi was responding to — and Commerce also reportedly asked the CFTC to freeze new compute derivatives for 60 days, putting CME's October 5 launch in doubt. Nothing signals a healthy market like a price nobody is allowed to publish.

    semafor.com
  3. 03

    Spain's data regulator logged the first known data breach carried out end-to-end by an AI agent.

    Someone pointed a well-known model at a Spanish organization and it chained the whole job itself: login, vulnerability search, modification of personal data, access to invoices, planning each next step as it went. The AEPD will not name the company, the sector, the number of people affected or the model, and the expert theories on offer are a jailbreak, a lab test that escaped its sandbox, or a pentester's homemade build — which are three different ways of saying nobody was watching.

    securityweek.com

So this week: the risk is a hoax, the price is classified, and the agents have started filing their own paperwork. Same time next week.

Never miss an episode

Subscribe to my AI podcast Promptly either on LinkedIn or Spotify. Last week in AI, in about two minutes, every week.