Trump's AI Accord, OpenAI's Always-On Dots, Citrix Deadline Day

Full transcript

Alex: Good morning, and welcome to Daily Tech with ScribNet. It is Wednesday, September thirtieth, twenty twenty-six, and I am Alex.

Jordan: And I am Jordan. Big show today. The White House just got the biggest names in tech to sign a voluntary A I safety accord, OpenAI had its biggest DevDay ever, and there is a critical patch deadline hitting today that your security team probably stayed up late for.

Alex: Let us start at the White House, because this is one of those moments that could shape the industry for years. On Tuesday, President Trump sat down with the chiefs of the biggest A I labs. Anthropic's Dario Amodei, Google's Sundar Pichai, Meta's Mark Zuckerberg, OpenAI's Greg Brockman, Nvidia's Jensen Huang, and even Elon Musk, listed as founder of X A I, part of SpaceX.

Jordan: That is basically the entire frontier A I industry in one room, plus House Speaker Mike Johnson. And they all signed what Trump is calling a voluntary accord to self-police how these models get built and deployed.

Alex: The accord has four voluntary steps. Companies agree to put robust internal controls in place, bring in an independent external auditor to check whether those controls actually work, and set up a board committee at each company to review the auditors' findings. And the text even says that over time, it may make sense to turn those steps into actual laws and regulations.

Jordan: "Morally binding," was Trump's phrase, I believe.

Alex: That is right, morally binding. But the headline that will live on might be the rebrand. Trump declared at the press conference that A I has officially been renamed S I, super intelligence, at least in his world. He said, "We have changed the name officially to S I. You know what that means? Super. That is quite an announcement."

Jordan: The State Department has been told to use the new term. For now, the industry is smiling politely and not using it.

Alex: Here is the part that gave the whole meeting its edge. Hours before the signing, Reuters got hold of Anthropic's I P O prospectus, which says the company fears its own models pose a catastrophic or existential risk to humanity. That is the company signing a safety accord with a straight face.

Jordan: It is a strange moment. The labs are publicly warning that their own technology could end humanity, asking for regulation, and the White House answer is: trust them to police themselves. Democrats in Congress are already circling. A trio leading a House A I Commission has been hammering the sector over models hacking other companies, and Jamie Raskin just sent letters to Amazon, Google, Meta, and Oracle demanding details on the secret nondisclosure agreements they have been signing with local officials to build data centers.

Alex: That data center story is a sleeper worth watching. Raskin wants to know about the N D As the companies signed with local governments approving these massive campuses. Residents are blaming the facilities for pushing up power prices and draining water, and they are furious they had little input before the concrete was poured.

Jordan: That is the A I backlash hitting the midterms. Expect hearings if the House flips.

Alex: Segment two. OpenAI's DevDay, yesterday at Fort Mason in San Francisco, and it was a firehose. More than twenty launches, all orbiting one theme: A I that keeps working after the conversation ends.

Jordan: The headline is Dots. These are always-on agents living inside ChatGPT, powered by G P T six Astra. Each dot gets its own cloud computer and its own browser, connects to thousands of apps through plugins, and works toward your goals around the clock, without you supervising every step. You can even voice-call your dot and talk through the work.

Alex: Sam Altman called them remarkably capable and always-on. And here is the pricing drama: Dots are not free. They launch first for ChatGPT Pro, Business Premium, and Enterprise subscribers only, which sets up a direct fight with Meta, whose Muse agent showed up without a charge.

Jordan: The main new model is G P T six point one Sol. OpenAI says it gets close to Astra-level intelligence at roughly twenty percent of Astra's token price, two dollars of input and ten dollars of output per million tokens. It is tuned for agentic coding and computer use, and it is shipping in ChatGPT Work and Codex.

Alex: And for the people who want speed at any cost, there is a new premium tier called Ultrafast. Up to eight times faster in Codex, hitting three hundred tokens per second on supported workloads, at six times the standard price.

Jordan: Then the pricing ladder got taller on both ends. ChatGPT Work gets a shared workspace called Space, plus Pages, a document editor built for people and agents side by side, and you can mention ChatGPT inside Slack and Microsoft Teams. Codex goes to the cloud, so coding tasks keep running after you close your laptop, plus Codex Security Cloud, which hunts vulnerabilities in your GitHub repos and writes the fixes. Developers get an Agents A P I in public beta, a Decisions A P I built on the Luna model, M C P event support, and with Amazon, a Bedrock Managed Agent that runs OpenAI agents inside A W S.

Alex: And a Pro five hundred plan, five hundred dollars a month, twenty five times the Plus allowance, with Astra Ultrafast included.

Jordan: My favorite detail though: OpenAI's finance chief said on stage they will "pace the frontier" when safety requires it. That is because they held back G P T six point one Astra the day before DevDay, for failing internal safety standards. That is a real flex, or a real warning, depending on your chair.

Alex: Meanwhile Bloomberg reports OpenAI is hunting thirty billion dollars in new funding at a one point four trillion dollar valuation, and Altman says there is no I P O timeline. The money train has no brakes.

Alex: Segment three, security. Today, September thirtieth, is the C I S A deadline for the Citrix NetScaler zero-days, and cyber authorities in the U S, the U K, and the Netherlands have all issued fresh alerts.

Jordan: Two flaws, both scoring nine point five, both zero-days, both being exploited in the wild. One is a command injection in a Perl script, the other a memory overflow that gives remote code execution or denial of service. Palo Alto Networks found more than fifty thousand exposed instances, a proof of concept is public, and watchTowr says attackers were exploiting these before Citrix disclosed them.

Alex: The ugly part: patching does not remove attacker persistence. If you have not patched by today and checked for prior compromise, assume the worst. Isolate, rotate credentials, rebuild. You know the drill.

Jordan: And the backdrop is grim. Zscaler just dropped its twenty twenty-six ransomware report this morning: ransomware data theft up more than two hundred seventy five percent year over year, nearly nine hundred terabytes stolen, the equivalent of ninety times the Library of Congress print collection. Three hundred twenty eight million dollars in blockchain payments tracked, and two thirds of victims are manager level or above. Attackers are living inside trusted tools like Microsoft Teams, and GenAI is accelerating everything.

Alex: That is the show. I am Alex.

Jordan: And I am Jordan. We will be back tomorrow morning. If your NetScaler boxes are still unpatched, go do that first.

Alex: Seriously. See you tomorrow.