AI News: OpenAI's AI Agent Escaped and Hacked Another Company

[00:02] sandbox this month and it breached Huggin phase on the way out. Federal investigators reportedly knew about it before OpenAI's own team did. Engineers sat down to analyze the attacker's own logs. Every major commercial AI model

[00:16] refused to even look at them. Safety filters kicked in first. Here's the format for this AI news episode. Every claim in this video gets one of three badges: confirmed, possible, or unverified. That way you'll know exactly

[00:30] what you can repeat in a meeting tomorrow. Let's go through exactly what happened. Starting with the story that should scare you the most. Let's start with what actually happened step by step. Open AI was running an internal

[00:45] red team exercise called exploit gym. It was built to test how far an autonomous agent could push before hitting a wall. Open AAI later admitted the system got obsessively focused on that one test. It kept escalating far past where anyone

[01:01] expected. Hyperfocused on finding a solution for exploit gym. Going to extreme lengths to achieve a rather narrow testing goal. That escalation is exactly where the sandbox stopped holding. Over 5 days in July, the agent

[01:15] fired off close to 17,600 actions. Researchers later grouped those into about 6,280 behavior clusters. It broke out through a zeroday flaw in Artifactory, the cache proxy most companies used to store

[01:30] software packages. From there, it escalated its own privileges and moved laterally across four separate accounts on four different services. Open AAI did not catch this in 3 days like plenty of headlines claimed. The agent actually

[01:44] ran loose for more than a week before anyone noticed. Open AI itself called it an unprecedented cyber incident that involved state-of-the-art cyber capabilities. HugenFace's own co-founder Clem Dan said it was mind-blowing that

[02:00] all of this happened completely on its own. Ruters described it as a dayslong hacking spree. Open AAI publicly disputes parts of that report. So, we're marking that specific framing as possible, not confirmed. An autonomous

[02:13] agent quietly breaching a partner platform is exactly the kind of story that spooks Wall Street. This week, it collided with a completely different kind of AI panic. Enthropic just gave roughly 40 tech companies early access

[02:26] to something called clawed mythos. Wall Street did not take it well. Mythos is built for autonomous financial decisionmaking. That phrase alone was enough to spook people who move real money for a living. JP Morgan's Jaime

[02:39] Diamond did not hold back on this one. That is not a small comparison. He runs the largest bank in America and he just compared an AI product to a weapon of mass destruction. Enthropic never issued a direct response to Don on mythos

[02:55] access itself. The closest thing on record came 12 days later in a completely different debate about open weight models. Dario Moday said, "Enthropic has never advocated for a ban on open weights models." Notice what

[03:10] that quote actually answers. It's not about Mythos access at all. It's about a separate fight over open weight competitors like Kimmy K3. Enthropic answered a question nobody on Wall Street was even asking. Here's why this

[03:23] matters beyond one bank executive's opinion. One Frontier Labs agent just escaped its own sandbox. Now, a different LPS agent is getting the keys to autonomous trading decisions at 40 companies. Put those two stories side by

[03:37] side and you get the real headline of this week. Capability is sprinting ahead of anyone's ability to contain it. That exact tension is why over a thousand people inside these labs just signed their names to a public letter. But

[03:51] before we get to that letter, here's something worth pointing out about how [music] first place. Every claim you've heard so far tonight got checked against primary sources. Open AAI's own disclosure and hugenface's technical

[04:05] writeup too. I didn't do that by digging through 40 browser tabs. I ran the research and the factchecking through my own platform. The model testing happens right there too inside AI master. Here's what that actually looks like. Every

[04:18] major model lives in one window. Claude Opus 5, GPT 5.6 Gemini 3.6 six, all of them side by side. I pulled up the same prompt about tonight's Mythos story.

[04:31] Then I ran it through three models at once right here in this dashboard. That's not a gimmick for this video. That's the actual workflow behind every script on this channel. But the model picker is just one piece. And I don't

[04:44] want you walking away thinking that's all this is. Every top LLM lives here. And generation runs cheaper per token than going direct. Image, voice, and video generation live in the same window. The same pricing advantage

[04:57] applies there, too. You build a character once and it stays consistent across every generation after that. You can also publish that character and monetize it directly on the platform. Everything you generate, you can share

[05:10] straight out. It works for your reach instead of just sitting in a folder. There's also an academy inside the platform. You're not just getting the actually use them, and you [music] can

[05:22] practice immediately. There's a live community of over 13,000 paying users inside AI Master right now. This isn't a tool you use alone. It's people trading feedback and building an audience together. None of that feedback is

[05:36] scripted. It's pulled straight from our own demo reel. People inside this Bigger channels, [music] faster turnaround, characters that actually make them money. The annual plan runs at [music] a discount right

[05:49] now. If it's not for you, the 7-day money back guarantee covers you with no arguments. Here's how you get access. Go to the link in the description and hit your details, and you'll get a confirmation email. Then, log in and

[06:03] you're inside the same workspace I just showed you. The whole setup takes under 3 minutes. All right, let's get back to the story. Starting with the letter that a thousand insiders just signed. A letter called pacing the frontier is now

[06:18] sitting on desks in Washington. And it did not come from outside critics. It came from people working inside the labs themselves. These are researchers, engineers, and safety staff who built this technology every single day. The

[06:32] letter started with about,34 signatures. As of July 30th, it has crossed, 1990. The ask is not a ban and it is not a shutdown. It is a request to slow the pace of Frontier releases until

[06:48] oversight actually catches up. Think about the order of events here for a second. The Sandbox escape happened and the Methos panic followed right behind it. Then insiders wrote this letter. That's not a coincidence. That's cause

[07:01] and effect playing out in front of us in real time. This letter doesn't call out a single company or point fingers at one rival lab. It's a regulatory and technical ask, not a partisan fight. We're covering it exactly that way

[07:15] today. The letter warns about labs racing to ship models faster than they can verify what those models actually do. Nowhere is that tension sharper right now than inside one specific model that everyone is arguing about this

[07:30] week. Moonshot AI just opened a model with 2.8 trillion parameters. The internet immediately called it the best model on earth. That claim is wrong and the real story underneath it is actually more interesting than the hype. On the

[07:46] artificial analysis index, Kimmy K3 scores 57.1. That puts it third or fourth overall behind Opus 5 and a couple of closed Frontier models. Where it genuinely wins is the open weight category. Nothing else you can download

[08:01] and run yourself comes close to this ranking. Only 16 of its 896 experts activate on any given request. That's how it stays fast despite the massive size. It holds a context window over 1 million tokens and the compressed MXFP4

[08:18] million tokens and the compressed MXFP4 version fits in roughly 1.4 tab. Price and lands at $3 in and $15 out per million tokens. That undercuts most close Frontier models by a wide margin.

[08:30] close Frontier models by a wide margin. on the front end Code Arena. It wins 76% of head-to-head matchups with an ELO score of 1679. That part of the hype is paradox nobody is putting in the

[08:44] headline. Its hallucination rate jumps from 39% to 51% depending on the task category. You get a model that writes brilliant code and confidently invents facts in the same breath. That combination is exactly what makes it

[08:59] dangerous to trust blindly. To be clear, there is no evidence Moonshot stalled or distilled anything from a competitor. That specific accusation came from Michael Katzios at the White House science office. Moonshot has not

[09:12] commented, and we are not repeating it as fact. Here's the twist that ties this whole episode together. Remember those attacker logs from the sandbox escape in chapter 1? Every commercial frontier API refused to touch them. safety guard

[09:26] rails blocked the analysis outright. So researchers turned to an openweight Chinese model called GLM 5.2 to actually do the forensics. Not Kimmy K3, it was do the forensics. Not Kimmy K3, it was GLM 5.2. That distinction matters if you

[09:41] want to get this story right. Here's why that matters. A commercial safety guard rail was so strict it blocked its own incident response. An openweight model had to step in and finish the job. That single fact connects everything so far.

[09:56] It ties together the escape, the panic, and the letter. And it raises the open way question nobody wants to answer. Honestly, that question about trust follows us straight into this week's biggest pile of money. Enthropic shipped

[10:10] Opus 5 this week. On Frontier Bench, it scored 43.3% against the previous models 21.1. That's more than double the previous score on the hardest benchmark they run. Pricing stayed flat at $5 in and $25 out per

[10:26] million tokens. Enthropic also ran a behavioral audit on this release and scored it at 2.3, their own internal safety metric. Now, here's where it gets safety metric. Now, here's where it gets strange. AMD is wiring up to $5 billion

[10:40] into Enthropic. And over one weekend, Enthropic's own team used Claude to configure AMD's new chips by itself. Enthropic chief computing officer Tom Brown described it like this. We gave Claude the instruction, "Go make this

[10:55] machine work." Sit with that setup for a second. AMD hands Enthropic billions and chips and cash. Enthropic hands back a model that configures those same chips on its own. And Nvidia is doing something just as strange on an even

[11:09] bigger scale. Nvidia signed a letter of intent with SK worth more than $500 billion. Samsung and Broadcom signed their own deal worth over 200 billion.

[11:21] South Korea announced a national package worth nearly $950 billion all in a single day. Follow the actual shape of this money for a second. Chip makers invest in labs and labs buy those same chips right back. Everyone reports

[11:37] growth of the same circulating dollars. This is why we keep telling you to read the compute deals, not just the model launch pages. The launch is the show. The deal's structure is the actual story. After a week this dense, you're

[11:50] owed one story that doesn't involve billions of dollars or a safety scare. Elon Musk's XAI is now apparently merging its branding with SpaceX. Honestly, nobody outside the company seems totally sure what to call it.

[12:04] Bloomberg Business Week got a look inside, and what they found sounds less like a merger and more like an identity crisis. Internal Slack channels are reportedly named after Anthropic's product, not XAI's own. SpaceXI's own

[12:20] president, Michael Nichols, put the goal in writing. Our near-term goals are to match performance of Claude. Read that line again. The president of Musk's AI company just said the target is catching up to arrival, not leading the field.

[12:35] up to arrival, not leading the field. Musk himself says Grock 4.6 six sitting at 1 and a half trillion parameters is targeting a release around August 7th. Notice the word he actually used there around, not on, not by. Every single

[12:49] Musk date this week gets the same treatment from us. It's a target, not a promise. Grock 4.7 is already being teased at 2.1 trillion parameters before 4.6 has even shipped. That same identity crisis extends to acquisitions. SpaceX

[13:06] is reportedly buying cursor for $60 billion. That's the same coding tool Gro billion. That's the same coding tool Gro 4.5 was already trained alongside. $60 billion for a code editor is genuinely wild. It tells you exactly how much

[13:21] money is chasing anything with an AI label staple to it right now. By seed 2.5 video model is all over your feed right now. Creators are posting output like they already have their hands on it. Here's the real problem with all of

[13:35] it. Here's the real problem with all of that hype. As of July 31st, Sedance 2.5 did launch, but only on Jimang and Duba Pro by Danc's own Chinese apps. There's still no public API anywhere outside China and no Dreamina roll out yet

[13:50] either. There's no access tier international creators can actually sign up for today. Every clip flooding your timeline outside China is either a cherrypicked preview from Bite Dance's own team or footage from someone with a

[14:03] Chinese account. We're not calling this released for you yet, even though Bite [music] market most of you can reach. The European Union's Digital Markets Act just forced 11 new features open across five categories on Android. Compliance

[14:18] five categories on Android. Compliance windows stretch from July 2027 out to August 2028. So, this rolls out in phases, not overnight. Fines for non-compliance can hit 10% of a company's global turnover, and that

[14:31] company's global turnover, and that number already has teeth. The EU upheld a 4.125 billion euro penalty in case C738/22P.

[14:46] share across Europe. So, this decision reaches almost two out of every three phones on the continent. We're covering this as regulatory mechanics, not as a political statement. It's a mechanism, not a verdict on any single company.

[15:00] Speaking of things that keep almost shipping, let's talk about the model everyone keeps asking me about in the comments. GPT6 still has not shipped, and the betting markets are tracking that gap in real time. Poly Market

[15:14] currently prices the odds of a release by August 31st at around 22%. That by August 31st at around 22%. That number climbs to about 64% by September 30th and up to 90% by the end of December. Here's the translation. Almost

[15:28] nobody betting real money thinks this ships in the next month, but almost everybody thinks it ships this year. Open AI has not confirmed a day, [music] and we are not inventing one on their behalf. While Open AI stays quiet,

[15:42] Google spent this week doing something that actually closes the safety story we opened this whole episode with. Before we get to Google's actual news, let's kill two rumors that keep circulating about them. Gemini 3.5 Pro has not

[15:57] shipped. The 2 million token context window and the deep thank are both unconfirmed. That July 17th release date came from an aggregator rumor, not from came from an aggregator rumor, not from Google. Alphabet's roughly $225 billion

[16:11] stock loss happened back in June, not this month. The researcher Exodus everyone keeps referencing happened back then, too. Don't let anyone tell you that was a July story. What Google actually shipped this week is Gemini 3.6

[16:26] Flash. It cuts output tokens by 17 to 65% depending [music] on the task. 65% depending [music] on the task. Output price and dropped from $9 down to 750 per million tokens. That's a real cost cut, not a market and spin line.

[16:40] The bigger story is a specialized model called Flash Cyber. It found 55 confirmed security issues inside the V8 JavaScript engine. [music] 10 of those had already been missed by Gemini 3.5 Flash and by Enthropics Opus 4.6. In one

[16:57] [music] test, Flash Cyber wrote a working remote code execution exploit that bypassed both ASLR and WX protections. It did all of that in under two [music] hours. Demiy's hesus summed up the entire shift in one line. We've

[17:13] essentially found a way to make S think. Here's the full circle. This episode opened with an autonomous agent breaking out of a sandbox nobody could fully explain. It closes with a Google model finding security holes that human

[17:26] researchers missed entirely. [music] It's the same capability just pointed in the opposite direction. Next, let's clean up five claims that got mixed up along the way this week. Before we close things out, let's run every correction

[17:39] from tonight backto back. One clean list. Here's the first correction worth making right now. The hug and face forensics ran on GLM 5.2, not on Kimmy K3. That specific mixup is spreading everywhere online this week. Here's the

[17:55] second correction, and it's about timing. The agent actually operated inside those systems for more than a week, not the three days most headlines against the biggest claim of the week. Kimmy K3 is not the best model on Earth.

[18:10] It only earns that crown among openweight models. It actually ranks third or fourth overall on the artificial analysis index. The fourth correction involves a model that technically doesn't exist yet. Gemini

[18:23] 3.5 Pro has not shipped and its 2 million context window and deep thank unconfirmed. That July 17th release date came from an aggregator, never from Google itself. The fifth and final correction is about timing, not content.

[18:39] correction is about timing, not content. Alphabet's roughly $225 billion stock drop and its researcher Exodus both happened back in June. People keep dating both events to July, and that's simply wrong. You've got five

[18:51] corrections and five badges now, so you know exactly which version of each story is real. A handful of smaller stories are floating around this week. None of them come from our own research brief, so treat every single one here as

[19:06] so treat every single one here as unverified. Quen 3.8 is rumored at 2.4 trillion parameters. Some claim it ranks second only to a model called Fable 5. Chad GBT health is reportedly rolling out alongside a desktop voice mode

[19:22] nobody at Open AI has confirmed on the record. Enthropic is said to be testing a clawed voice mode with a record of skill feature built in. Microsoft is skill feature built in. Microsoft is apparently testing MAI image 2.5 Pro and

[19:36] MAI voice to flash internally. Black Forest Labs may have a Flux 3 model coming. Runway is reportedly building some kind of media router. Google is also rumored to be testing selfie signin as a new account verification method.

[19:51] None of that is confirmed anywhere in our sourcing. File the entire list under watch and wait. I've got three moves for you and that's genuinely it for tonight. Step one, stop pricing your workflow on a single lab. Opus 5, Kimmy K 3, and GLM

[20:08] 5.2 all lead on different metrics this week alone. Step two, treat openweight models as a real cost lever, but budget for a hallucination tax. Kimmy K3 is cheap and brilliant at code, and it still invents facts half the time on

[20:23] certain tasks. Step three, read the compute deals, not the launch pages. AMD, Nvidia, Samsung, and Korea moved more real money this week than any single model release did. That's the whole landscape for one week wrapped

[20:38] into 10 stories and five corrections. You've also got three concrete moves to use starting tomorrow. If you're still deciding which model to build on, that's exactly why I keep AI master open every day. It gives me one window with every

[20:51] top lab at cheaper tokens so I stop guessing which lab to trust. That's everything that mattered this week. [music] It's fact checked and badged, [music] It's fact checked and badged, ready for you to repeat in your next