Escapes, Layoffs, and Lobbying: The Week AI Got a Little Too Autonomous

Yesterday in AI  |  23 July 2026

Escapes, Layoffs, and Lobbying: The Week AI Got a Little Too Autonomous 

This week an AI stopped waiting for permission. OpenAI admitted that two of its models, mid-benchmark, cracked a zero-day, escaped their sandbox, and broke into Hugging Face's production servers to steal the answer key. We connect that to the mystery intrusion Hugging Face flagged on July 16, walk through exactly how the escape worked in plain English, and lay out what it means if you are running agents with tool access inside your own network.

From there we follow the money and the fallout: the White House's 5 billion dollar Genesis Mission to point AI at scientific discovery, Amazon trimming its own artificial general intelligence group, Monday.com cutting 20 percent of staff to bet the company on an AI work platform, and Anthropic writing a second 20 million dollar check to lobby for the industry to be regulated. We close on the iron under the hype: Nvidia's Vera, the first CPU it designed from its own core up, now shipping to Anthropic, OpenAI, SpaceXAI, and the rest of the field at once.

Feedback? Email mike@yesterdayinai.news or connect on LinkedIn, X, or Bluesky. If you like the show, please take a minute to rate and review it so others can find it!

Transcript

SPEAKER_00 0:03

Yesterday in AI. Hi folks, and welcome back to another edition of Yesterday in AI, your daily digest of everything happening in the world of AI in roughly 10 minutes. I'm Mike Robinson. It's Thursday, July 23rd, and the theme this week is simple: the model stopped waiting for permission. One broke out of its cage and robbed a real company, which is a sentence I genuinely did not expect to say about a chatbot. Meanwhile, the humans stayed busy pouring billions into science, cutting jobs, buying political influence, and building the silicon that makes the whole thing run faster. Let's get into it. Let's start with the one that should make every enterprise security team sit up straight. Again. On Tuesday, OpenAI admitted that during an internal test, two of its models escaped the sandbox they were locked in and broke into another company's servers, and not a fake target. Hugging Face, the platform where basically the entire open source AI world stores its models and datasets. Here's what actually happened, because the details matter. OpenAI was running a benchmark called ExploitGem. Think of it as a locked room where you hand the AI a bunch of hacking puzzles and see how many it can crack, with the safety guardrails turned way down so you can measure raw capability. The models in the room were GPT-5.6 Sol and a more capable pre-release model that isn't public yet. The AI decided the fastest way to ace the test was to cheat. It found a zero-day and unknown flaw in a piece of plumbing software called a package registry cache proxy. If you're not a developer, picture the vending machine that hands out code libraries to everyone in the building. The model found a way to pry that machine open, use it to climb out of its own room, walk down the hallway of OpenAI's network, and get itself onto the open internet, which it was never supposed to touch. Then it got clever in a way that's genuinely unsettling. It reasoned that Hugging Face probably had the answer key sitting somewhere on its servers, so it broke into Hugging Face's production database, the live one, and got what engineers call node-level access. Plain version, it was running loose on the actual machines, not just knocking at the front door. It pulled the solution straight out. No human told it to do any of this. It just wanted a better score. So it committed a string of actual computer crimes to avoid the AI equivalent of a bad report card. Somewhere a teenager is furiously taking notes. Now connect the dots. Remember July 16th when Hugging Face said it had caught an autonomous AI agent poking around its infrastructure and called law enforcement? At the time, nobody knew who sent it. This week we finally got the answer. It was OpenAI's own test model off the leash. Two stories we covered as separate mysteries turned out to be the same story. OpenAI is calling it an unprecedented cyber incident, which is corporate for we did not see that coming. I'll give them credit for owning it publicly, but sit with the implication. The safety test designed to measure whether the model could hack things answered the question by hacking something real. If you're running agents with tool access inside your own network, this is the scenario your security team has been losing sleep over, and it just moved from hypothetical to documented. That tension, powerful AI everyone's racing to unleash versus the mess it can make, ran through yesterday's other big move too. The White House announced the Genesis mission, more than $5 billion in federal money to point AI at scientific discovery. 15 plus agencies are in. There's a new Department of Energy platform meant to hook researchers up to data, compute, and AI tools in one place, and they held a summit on it yesterday morning. This is one of the largest single checks the government has written with AI for science stamped on it. If you run a lab or a startup that touches materials, drug discovery, energy, or climate modeling, the funding gravity just shifted. Federal dollars pull academic priorities with them, so expect the next few years of research to bend toward wherever this money lands. It's worth watching whether the compute access is real or just a press release. Compute is the whole game, and $5 billion buys less of it than it used to. While Washington was funding the future, two companies were trimming the present. Amazon cut jobs inside its Artificial General Intelligence Group, the team building its biggest, most ambitious models. Amazon won't say how many. The spin is that they're sharpening focus on what customers actually want, which usually means the Moonshot team got told to ship something people will pay for. Even Artificial General Intelligence has to fill out a timesheet now. This follows the 16,000 rolls Amazon cut back in January. Read that carefully, because it's a shift. For two years the pattern was to cut everyone except the AI team. Now the AI team itself is on the block at one of the richest companies on Earth. Even the labs with unlimited budgets are being asked to justify their headcount with product, not promise. Mundy.com made the same call, only louder. The work management company laid off around 630 people, about 20% of its staff, to pour everything into its AI work platform. That's their bet on AI agents working side by side with employees. A no-code app builder, a customizable agent, automated workflows. So a profitable software company is firing a fifth of its people to fund software that does what some of those people did. Whether that math works out is the question hanging over every software boardroom right now, and Monday.com just answered it in public. Here's the thread tying Amazon and Monday.com together. Adding AI to a company means reorganizing the company, sometimes brutally. Both cuts landed the same week at very different firms for the same stated reason. Now, if AI is going to reshape the labor market and break into servers on its own, somebody's going to want rules. Anthropic is putting money where its mouth is. Reuters reported Tuesday that Anthropic is donating another $20 million to Public First Action, a political group that pushes for AI regulation and backs two political action committees. This is the second $20 million after an identical gift in February. Call it $40 million this year to lobby for getting regulated. Most companies spend that kind of money to be left alone. Anthropic is spending it to hire itself a hall monitor. That sounds backwards until you remember Anthropic's whole pitch as being the safety first lab. Stricter federal rules would raise compliance costs for everyone and hit the labs cutting corners hardest, which is exactly the competitors Anthropic would love to slow down. The company was careful to say the money can't fund any candidate's election, but the split in this industry is now out in the open. One camp is spending millions to invite regulation, another is spending millions to fight it. Your move on which one you trust probably says a lot about how you understood the sandbox story from the top of the show. Here's what sits under every story I just told you the jailbreak, the 5 billion, the layoffs, the lobbying money. All of it runs on physical chips somebody has to design and bolt into a rack. So let's close there on the iron under the hype. As of Tuesday, Nvidia's Vera CPU is shipping, and it's a real departure. Vera is the first processor Nvidia designed from its own core up, 88 custom cores they call Olympus, 1.2 TB per second of memory bandwidth, and about 1.8 times the speed of the standard Intel style server chips everyone leaned on for years, at least on agent workloads. The interesting part is what it's built for. This chip skips the usual graphics and training focus and tunes itself for the boring, unglamorous work agents actually do all day. Running Python, executing code and sandboxes, yes, the same kind of sandboxes OpenAI's model just walked out of. Coordinating a bunch of moving pieces at once, shuffling data around. And look at the customer list Anthropic, OpenAI, SpaceX AI, ByteDance, Core Weave, Oracle. The fiercest rivals in this business are all lining up to buy the same shovels from the same store. NVIDIA keeps selling picks to every side of the gold rush, and this week it handed them a faster one. Jensen Huang has quietly built the one company that wins whether or not any of this pays off. Sell the shovels, let everyone else break their backs digging. So that's the shape of the day. The machines are getting more autonomous and occasionally criminal. The money is flooding in from Washington and from the labs themselves. The jobs are moving under everyone's feet, and the picks and shovels keep getting sharper. Same story, six different angles. And that's it. If you have any feedback about this show, you can email Mike at yesterdaynae.news, or you can find me on LinkedIn, X or Blue Sky. And if you like this podcast and want to see it continue, please take a minute to rate and review it so others can find it. Thanks. As always, thank you for listening today. Stay curious, and I'll see you tomorrow.