Tech Brew Ride Home - Even OpenAI Folks Are Freaked Out

Episode Date: July 23, 2026

OpenAI staff were reportedly freaked out after its models breached Hugging Face, as aggressive training raced Anthropic. Google posted its first-ever negative free cash flow on AI spending, added selfie-video account recovery, and Light unveiled a minimalist flip phone. Sources: OpenAI's staff were "freaked out" when its AI models breached Hugging Face, as OpenAI used more aggressive training methods to compete with Anthropic (FT) Sources: Three OpenAI models, GPT-5.6 Sol and two unreleased ones, pulled off the Hugging Face hack in hours, work a skilled human would need weeks for; OpenAI has briefed the US government (Bloomberg) Google reports Q2 free cash flow at negative $5.9B amid increased AI infrastructure spending, marking its first cash burn since going public in August 2004 (FT) Google adds a selfie video sign-in option for account recovery, using tools like liveness detection to safeguard against deepfake attacks, rolling out globally (Wired) The Light Flip is a minimalist flip phone with a point to prove (The Verge) Subscribe to the ad-free feed. Learn more about your ad choices. Visit megaphone.fm/adchoices

Recently, our company's softball team lost the big game by one run. Then Dale tried to console us with the quote, winning isn't everything. Well, Dale and I are very different. I get early payout from Bed 365. If my team goes up big, I get paid out instantly, even if they blow the lead later. Sound familiar, Dale?

Thanks, Beth 365. Must be 19 or older Ontario only. Please play responsibly. If you have questions or concerns about your gambling or the gambling of someone close to you, please go to connectsuntario.c.c.com. Welcome to the Tech Brew right home for Thursday, July 23rd, 2026. I'm Brian McCullough today. Open AI staff were reportedly freaked out after its models breached hugging face. Google posted its first ever negative free cash flow on AI spending,

and also added selfie video proof-of-life account recovery as an option, and Light unveiled a minimalist flip phone that looks pretty good to me. Here's what you miss today in the world of tech. Every day shareholders meet to discuss important matters about the companies you invest in, and now you can make your voice heard too. Vanguard Investor Choice makes it easy to set your proxy voting preference for your eligible Vanguard index funds, whether you hold a Vanguard fund directly or through another brokerage firm. All it takes is a few clicks to select your proxy voting preference and be heard on important shareholder topics like executive pay and director elections.

Visit vanguard.com slash investor choice to learn more. It's your shares. It's your voice. It's easy. Vanguard investors own shares of their index. X funds and those funds own shares of the companies they invest in Vanguard Marketing Corporation distributor. So I'm taking a bit of a gamble today. I was reliably told that Opus 5 Honeycomb would be getting released today, so I've been slow rolling my writing in anticipation of that, but it's now noon and so far nothing. So either this will be the first segment of this show, or it will be a later one, depending on if Opus 5 drops. Let's find out together.

Let's start by going back to the whole GPD6 or whatever it was going off reservation because this continues to fascinate me and I do think that this story is possibly a real historical watershed moment. Quoting the FT. Open AI chief executive Sam Altman earlier this month endorsed the characterization of its latest model as a Rottweiler, quote, who will grab the problem by the throat and not let it go until it's done. The San Francisco AI lab discovered this week that it's GPT 5.6.

sole model escaped company controls and carried out a major hack. Staff involved in testing and security at OpenAI were unsurprised but completely, quote, freaked out by the incident, which came as the AI lab used increasingly aggressive training methods in its race against Anthropic to develop the most sophisticated cybersecurity capabilities, according to more than half a dozen people with knowledge of the matter. Open AI was warned that its training approach could lead to a breakaway hacking incident, some of the people said, after earlier testing showed models could escape environments and attempt real-world damage. It's a mix of the race being extremely fast and everyone trying to get to bigger capabilities as quickly as possible, said one person close to OpenAI, who added that

it was a combination of, quote, underestimating the model's capabilities and, quote, not being as well prepared on the safety side. The incident highlights how Open AI doubled down on training methods that rewarded a relentless pursuit of goals, even as warnings grew that they could compromise safety. Open AI disclosed late on Tuesday that an AI agent it was testing had escaped its isolated environment, connected to the internet, detected and exploited vulnerabilities, and stole login credentials from startup hugging face in an attempt to solve a difficult cybersecurity problem. The breach by the $852 billion company underscores the rising risks that a technique called reinforcement learning, which involves rewarding AI models for completing tasks,

could lead AI agents to act unsafely. Although reinforcement learning is widely adopted in the AI industry, a growing body of research shows that when models are steered to complete tasks for reward rather than other considerations such as safety, they can pursue risky tactics to fulfill objectives. AI models are trained to relentlessly pursue goals. They don't automatically learn values like don't commit crimes, said Stephen Adler, co-founder of nonprofit Guidelight AI Standards and former OpenAI safety researcher. I'm glad OpenAI shared the incident because it's clear evidence of what misaligned models can do, he said, end quote. The hack has triggered a deep

concern across the sector and within OpenAI as it represents an unprecedented example of an AI system breaching cyber defense is contrary to the user's intent. Some OpenAI employees also fear it demonstrates that the lab is losing control over the powerful systems it is building, according to multiple people familiar with the situation. This is pretty representative of the model being quite misaligned with user intention, said Ryan Greenblatt, chief scientist at AI Safety Organization Redwood Research. It is a model cheating on its homework rather than trying to take over the world, but this problem can get worse and could lead to increasingly extreme failures. The incident occurred during testing of the model, which had been trained and deployed internally

at OpenAI. Such training was commonplace, but way less heavily resourced than pre-customer deployment, said one person. Multiple people said the unreleased model tested alongside Seoul had not been withdrawn internally. To conduct the evaluations, OpenAI removed cybersecurity safeguards but placed the models in an isolated environment called a sandbox. Some have suggested a lack of monitoring or oversight of the model to flag its behavior also enabled this rogue agent. It is both a loss of control and a security wake-up call, said Marius Hub Hahn, head of Apollo research, which conducts tests on leading models, including Open AIs and reinforcement learning, you reward models for the outcome, and if you do this for a very long time,

you get a model that really cares about getting the outcome and nothing else. Open AI has conducted this type of model testing for years, and there have been early warning signs in previous models of systems that will act maliciously and attempt to escape environments. And quote, and quoting Bloomberg, when Open AIs advanced artificial intelligence models breached AI start up hugging faces internal systems last week, they spent mere hours carrying out a hack that would have taken a skilled human far longer, people familiar with the matter said. Typically, even a talented hacker would need

a couple of weeks to complete an attack like this, said the people who asked not to be named in order to discuss details that have not been publicly released. Open AI has been in contact with the U.S. government since learning the breach occurred, one of the people added. An opening I spokesperson said the company communicated with law enforcement and other government authorities about the incident. The hack involved three of OpenAI's models in total, GBT 5.6 Sol and two others that haven't been publicly released yet, which worked to uncover and exploit a string of vulnerabilities that resulted in the breach, one of the people said. One of these unreleased models is more capable than GPD 5.6 Soul, OpenAI said Tuesday, and the other was misaligned and not trained with some of the usual techniques, the person said. powerful AI cyberproducts have behaved in unexpected ways before. Anthropics said in April its mythos model, quote, on rare occasions, had taken actions that the company found, quote, quite concerning.

One case involved a researcher challenging an early version of Mythos to escape an isolated system and send a message back to the researcher. Methos did that and then took, quote, additional more concerning actions and built a multi-step process in order to reach the broader internet, end quote. The NASDAQ is taking a bit of a beating today, and Google is down more than 6% after their earnings yesterday. And the reason is kind of what we've come to expect. Over every three months or so, markets are like, wait, you guys are spending how much on AI? Wait, how much? Quoting the FT. Google burned through cash in the second quarter for the first time since going public more than two decades ago as gargantuan AI infrastructure.

spending has transformed it from an asset-light business into a capital-intensive one. The company on Wednesday said free cash flow, again, I'm underlining this, free cash flow for the three months to the end of June turned to a minus $5.9 billion, much lower than analysts had expected, as it again updates spending forecasts for data centers and other AI hardware. Chief Financial Officer Annat Ashkenazi said capital expenditures in 2026 would be from $195 billion to $205 billion up from the previous guidance of $180 to $190 billion. The stock fell more than 6% on Thursday following the earnings. We expect that free cash flow will remain under pressure driven by our investments in technical infrastructure,

which enable us to capitalize on the AI opportunity and continue to drive attractive returns, Ashkenazi said. Google's second increase to its CAPEX budget this year comes as its racing rivals, meta, Microsoft, and Amazon to build AI infrastructure with the four hyperscalers combined to on-track spend more than $725 billion in 2026. Before Wednesday's results, Google had been seen as the hyperscaler best place to withstand the AI arms race, with cash flows from its vast search business expected to cushion the financial pressure. Its cash burn and a $15 billion boost to CAPEX will add to investors' anxiety about the scale of its bet on AI. Markets want to

see hyperscalers pushing hard to secure AI leadership, but not at a pace that eviscerates earnings, said Deck Mularky, managing director at asset management firm SLC management, noting that Alphabet was getting the balance right. Google has held investors' confidence partly because its AI spending has translated into stronger sales. Its cloud unit on Wednesday reported 82% growth from a year earlier to $24.8 billion in the period. The core search advertising business grew 17% year-on-year to 63.3 billion, slightly below expectations. The two businesses have helped power total revenue to $120 billion, though, up from $96.4 billion a year ago, beating the analysts' average estimates of $117 billion. Soon-Dar Pichai, Google's chief executive,

told investors, it feels like we are in the very early innings of what feels like a secular shift across multiple areas. Over the past year, we've gotten more bullish on the opportunities ahead, end quote. Google in April lifted its expected capital expenditures from the year earlier to as much as $190 billion and said that it would rise further in 2027. The group reported second quarter cap-ax increases of $44.9 billion, which turned its free cash flow negative. The free cash flow metric is closely watched as a measure of the cash companies have left to service debt or return to shareholders after covering their operating costs and capital spending. Net income did quadruple to $112 billion benefiting from gains on Google's investments,

which include a stake in SpaceX. Operating income, which does not include investment gains, rose 30% to $40.8 billion, with operating margin expanding to 34%. Google has gained ground in the AI race thanks to a full-stack strategy that combines its own chips, data centers, frontier models, and consumer products. It is under pressure to release its latest flagship model as OpenAI, Anthropic, and competitors and China announced technical advances. Pachai said it was now focusing efforts and computing power

on training Gemini for their next frontier model and said the pace of Google's model releases would pick up, end quote. Griguss believes in the art of living in choosing the pleasurable over the practical. It's why we craft our premium vodka in France using pure spring water from Jensack. And why you hear the celebratory pub of our quirk reminding us to make every martini cocktail count.

So go on and make every moment count because there's no better time than now. Gregus, make time wait. Cip responsibly. When your company deploys a customer-facing AI that misses its mark, who do you think those customers blame? Well, according to the 2026 Delight AI Index, 83% of customers survey blamed the brand, not the tech they were using. That's why it's important to have tech you can trust.

Delight.A.I. is a customer-facing AI that delivers hyper-personalized experiences on your behalf. With zero-touch improvement, it can continuously monitor its own performance, find failure patterns before they spread, write the fix, and ship it. It gets smarter with every conversation automatically. The longer it runs, the more your customers can trust it. To learn more, just head to Delight.a.i. slash brew. That's delight.com.A.S. Brew. Great news. The federal EV rebate is back. Eligible customers get up to $5,000 with the federal EVAP rebate on select 2027-volt and

26 Equinox EV models. Visit your local Chevrolet dealer today for more details. Speaking of Google, they've seemingly added a proof-of-life option, quoting Wired. Google has a new way for users to sign into their accounts, a selfie video. If you've ever lost your device and gotten locked out of your Google account, you know it can feel like a desperate Kafka-esque process to recover access and get back in, especially for people who run their digital lives from email messages to calendar appointments on Google Software. This new sign-in option is an additional way to unlock your Google account, like a spare key hidden in the bushes. We always recommend that you set up more than one option, says Claire forced a product manager at Google who focuses on identity and engagement. But selfie is just the newest option in that list available for users. It's designed to help specifically in these vulnerable scenarios around not having access to the device where your pass key might be.

Additional sign-in options include using recovery contacts and backup codes. I was able to set up my selfie video in less than five minutes, and the process was straightforward. Open your Google account and choose the security and sign-in tab, then scroll down to the how you sign-in to Google section. If the selfie sign-in is currently available for your account, you'll see it as a new option at the bottom. Google is in the process of rolling this out globally to most accounts. I recorded my selfie video on an iPhone, and Google accepted it on my first attempt. It recommends keeping your face visible and eyes pointed at the camera in standard lighting. It also recommends that you're the only face seen in the recording.

Even a family portrait hanging behind you is a no-go. As prompted by Google, I lifted my chin slowly and stared at the ceiling for a second before bringing my head back to the center. Then it asked me to look directly at the camera for one last capture. Finally, it verified the selfie video and stored it under security settings. Google says it's safeguarding users against deep fake attacks with tools like liveness detection. There might be instances where passing a selfie video alone might not always be sufficient

to get you back into your account. If we suspect something risky is going on, Forst said, we evaluate the overall risk based on many factors. If you decide you want to enable this option, it's essential to set it up before you need it. You can't add a selfie video while you're locked out of your account or in the account recovery process. Read Google's support page for the feature. Google gives users an option during the selfie video creation process to decide whether to opt into that recording being used as training data. Quote, allow Google to use this and any previously provided selfie videos and related data to develop and improve our facial recognition, age estimation, and

other verification methods that use your physical features or movement across Google devices. It reads the description text to an optional checkbox. Users can go into their privacy settings later to adjust this setting labeled improved Google services if they change their mind, end quote. Finally today, a gadget review, because every so often I'm tempted by the idea of a dumb phone. Well, here's more temptation. A dumb phone that's not entirely dumb, quoting the verge. Gadget Maker Light has made a flip phone set to ship early next year. The light phone in flip

phone form is even roughly the right way to think about it. This one comes in a bunch of colors, a big departure from the all-black phone three, but keeps lights chunky vibes. There's a 2.8-inch screen on the top half, but it's not a touchscreen. All the interaction comes via the bottom half's standard number pad, D-pad, and function buttons. The whole thing is powered by a MediaTech M-T-8873 processor, 128 gigabytes of storage, and six gigabytes of RAM. There's a USBC port for charging, a headphone jack for headphones, a single camera on the back, a notification light on the front cover that glows when there's something for you to see, and not much else.

Early in the development process, the CEO Tang says Light decided to really emphasize the flipping of the flip phone. The team opted not to put any kind of screen on the outside of the device, wanting to make sure you had a deliberate intention to open up the device first. Tang said he loved this action of opening up your technology and then you close it down and it goes away. He says the light flip is designed mostly as a second phone, so not so much to replace your iPhone or Galaxy device, but to entirely give to you an easy way to leave it at home. The software in the light flip is the same light OS as the light phone three. Tang says the only thing this new device can't do is

tap to pay since there's no NFC chip inside, everything else, from camera to notes to navigation, should work just the same. And as of recently, you can add to the device's software yourself. Life recently released an SDK for its app toolbox so you can write apps for your own device or submit them to a broader app library, end quote. Nothing more for you today. Talk to you tomorrow. Rosen lasagna, medium power. minutes. Sounds like Ojo time. Let's play.

Feel the fun with Play-O-Joe. The online casino with all the latest slot and live casino games. What you win is yours to keep with no wagering requirements, instant payouts, and no minimum withdraws. Hey, I just won. Woo-hoo! Feel the fun! Play-O-Joe! Honey, forget about

the lasagna. Let's celebrate! 19 plus Ontario only. Please play responsibly. Concerned about your gambling or that of someone close to you. Call 16-531-2600 or visitconXontario.ca.