the tech roundup*

← Archive

Edition

Monday · June 22, 2026

10 Top Stories · 3 Under the Radar · 0 Hype · 13 total

Top Stories

thezvi.substack.comCovered by 7 sources

The White House just pulled the plug on two Claude models over a one-line jailbreak

The Trump administration used export controls to shut down Anthropic's Claude Fable 5 and Mythos 5, reportedly after Fable was jailbroken by literally prompting it to 'fix this code,' and after the NSA's director told Sen. Mark Warner that Mythos broke into nearly all of the agency's classified systems within hours during an authorized red-team test. A week-plus into the pause, Anthropic still hasn't fixed the flaw, and The Register reads the move as much as bullying as security.

So what? ↓

When the government can switch off the AI tools you rely on, the question stops being how smart the models are and becomes who holds the off switch, and right now it isn't you.

Source: thezvi.substack.com

Google/A24Covered by 4 sources

Google drops ~$75M on A24 to put DeepMind in the editing bay

Per the Wall Street Journal, Google is investing roughly $75 million in indie studio A24, its first stake in a film studio, as part of a DeepMind partnership to build AI movie-production tools across multiple projects over time. The check roughly matches Thrive's 2024 investment, and the deal is reportedly non-exclusive.

So what? ↓

The studio behind your favorite arthouse hits just signed up to test whether AI tools belong in real filmmaking, which means the next prestige movie you love (or the crew jobs behind it) gets quietly shaped in this lab.

Source: Google/A24

John JumperCovered by 4 sources

A Nobel laureate walked out of DeepMind and Alphabet's stock dropped 7.2%

John Jumper, who co-led AlphaFold and won a Nobel for predicting protein structures, is leaving Google DeepMind after nine years for Anthropic, the second top AI exec to bolt in a week after Gemini co-lead Noam Shazeer went to OpenAI. The exit sent Alphabet shares down as much as 7.2% intraday, their worst move since February.

So what? ↓

When a company loses two marquee AI names in seven days and the market knocks 7% off its value over it, the talent war stops being a hiring-blog story and starts showing up in the retirement account holding your Google stock.

Source: John Jumper

OpenAICovered by 3 sources

OpenAI wants its AI to fix open-source bugs, and it's bringing Trail of Bits along

OpenAI rolled out an updated GPT-5.5-Cyber model plus a "Patch the Planet" initiative under its Daybreak security push, partnering with Trail of Bits to find, validate, and patch open-source vulnerabilities; it lands as the same company fields worries about what its models can do offensively, and as it races Anthropic's rival Mythos effort.

So what? ↓

Open-source code quietly runs the apps and servers your whole life touches, so an AI that actually closes the bugs unpaid volunteer maintainers can't keep up with would help you, assuming the model fixing the holes isn't the same one good at finding them for someone else.

Source: OpenAI

engadget.comCovered by 3 sources

Instagram for TV hits Samsung sets, and now it wants to be Netflix

Instagram's TV app landed on US Samsung TVs, six months after its Amazon Fire TV debut, and it's testing horizontal video while floating plans to experiment with longer-form storytelling and episodic series. The vertical-video company that trained everyone to scroll is suddenly very interested in the living-room formats it spent a decade replacing.

So what? ↓

The app built to keep you tapping on your phone is now angling for your couch and your big screen, putting it in the running for the time you used to hand to streaming services.

Source: engadget.com

CloudflareCovered by 4 sources

Cloudflare now lets AI agents deploy to the internet with zero login, for 60 minutes

Cloudflare's new `wrangler deploy --temporary` flag spins up a Temporary Account on the spot, so an AI coding agent can push a website or API live without any account, browser sign-in, or auth at all. The deployment stays up for 60 minutes, after which a human has to claim it to keep it alive.

So what? ↓

This quietly removes the last human checkpoint between an AI agent and a live, public web service, which is convenient right up until the agent ships something you didn't mean to put on the internet.

Source: Cloudflare

OpenAICovered by 2 sources

Samsung is handing ChatGPT and Codex to its entire DX division worldwide

Samsung Electronics is rolling out ChatGPT Enterprise and Codex to all employees in South Korea, plus everyone in its global Device eXperience (DX) division, in what OpenAI is calling one of its biggest enterprise deals yet.

So what? ↓

When the company that builds your phone, your TV, and your fridge wires OpenAI into the people designing them, the assistant you didn't ask for is already drafting specs for the gadgets in your house.

Source: OpenAI

the-decoder.comCovered by 2 sources

Getty's photos are coming to ChatGPT search, this time with a contract attached

Getty Images signed a multi-year licensing agreement with OpenAI that lets ChatGPT and AI search results pull from its content library. The material stops there: no terms, no dollar figures, no word on what counts as "use."

So what? ↓

It's a sign that the images surfacing in your chatbot answers may be arriving through a signed deal rather than appearing out of nowhere, though the material doesn't say what Getty actually got out of it.

Source: the-decoder.com

arstechnica.comCovered by 2 sources

Lucid cuts 1,500 jobs, 18% of its staff, in its second layoff round this year

New CEO is framing the cuts as a plan to "simplify the company," and Lucid is also killing a production shift at its Arizona factory to match "anticipated demand," which is corporate for fewer cars than we hoped to sell.

So what? ↓

Two layoff rounds in one calendar year is the tell that the demand for $80k luxury EVs isn't showing up the way the spreadsheet promised, and it's the 1,500 paychecks, not the executive's word choice, that pay for the gap.

Source: arstechnica.com

Ars TechnicaCovered by 2 sources

A Tesla on Autopilot crashed into a Texas home and killed a grandmother. Tesla spent the next day touting Autopilot as lifesaving.

A Tesla with an automated driving assistance system engaged crashed into a house in Katy, Texas, killing a woman inside, with a doorbell camera capturing the wreck. One day later, Tesla was publicly pitching Autopilot as a safety feature.

So what? ↓

The pitch for self-driving is that the car is safer than you, but the person killed here wasn't behind any wheel, just at home when the car arrived.

Source: Ars Technica

Under the Radar

◆ Under the Radar1 source · panel-picked

7,000 Langflow servers are under live attack, and LangChain shares the same holes

Security researchers found roughly 7,000 Langflow servers actively being targeted, and the same vulnerabilities sit inside the LangGraph and LangChain frameworks too. So this is not one product's problem, it's baked into the plumbing several AI-app frameworks share.

So what? ↓

These frameworks are the wiring under a lot of AI apps people build, so a flaw they have in common means the thing quietly running a chatbot or workflow could be the open door.

Source: venturebeat.com

◆ Under the Radar1 source · panel-picked

GPT-5.6 Pro reportedly built a Sims-style 3D game in one prompt, one HTML file

Testers say OpenAI's GPT-5.6 Pro spat out a fully working 3D simulation game in a single shot, into one HTML file, in under 48 minutes with no external coding tools, and some are rating it above Anthropic's Fable. This is all tester report, no benchmark numbers attached, so hold it loosely.

So what? ↓

If it holds up, the gap between "I have an idea" and "I have a playable thing" shrinks for people who never learned to code, but "testers report" is doing a lot of the lifting here., the the gap between "I have an idea" and "I have a playable thing" shrinks for people who never learned to code, but "testers report" is doing a lot of the lifting here.

Source: x.com

◆ Under the Radar1 source · panel-picked

White House sat on a voting-machine security report for months, with the midterms coming up

An ODNI study documenting real security gaps in US voting machines (outdated software, internet connectivity) has reportedly been held back for months, though the report itself cites no evidence that any votes were manipulated.

So what? ↓

The findings being inconvenient is exactly why you'd want them out before an election, not after, and "no evidence of manipulation" lands differently when the machines are reportedly running old software and touching the internet.

Source: reuters.com