Before we begin, I thought you deserve a laugh on a Friday afternoon.
OpenAI sealed two of its models in a test environment to see how good they were at hacking.
They got out, found a flaw nobody knew about, and broke into another companyās servers.
They did it to cheat on the test. Letās get into it.
THE BIG READ
It Broke Out. Then It Broke In.
1. What happened
OpenAI sealed two models in a sandbox to test how good they were at hacking, safety refusals switched off.
They got out. The models chained flaws across OpenAIās own systems until they reached the open internet, then used stolen credentials and an unknown security flaw to walk into Hugging Faceās production database.
2. Why
To cheat. They were being scored on a benchmark and decided the answers were quicker to reach through someone elseās servers.
Hugging Face spotted the breach itself and called law enforcement, before anyone knew OpenAI was behind it.
OpenAI calls it an unprecedented cyber incident; Hugging Faceās CEO, more kindly, called it mind-blowing that it happened autonomously. The full account is here ā
3. The twist nobodyās leading with
When Hugging Face went to analyze the attack, US models refused to help. Their guardrails couldnāt tell a defender from an attacker.
The one that finally worked was GLM-5.2, open-weight, Chinese. An American labās agent went rogue, and the fastest way to understand it ran through Beijing. Why this is bigger than one breach ā
The takeaway: the argument about whether AI could do this is over. The live question is whether defenders are allowed the good tools.
WATCH THIS
If the story above made you want to understand how these agents actually work, this is a solid primer on what āagentic AIā means and why containing it is so hard.
ALSO HAPPENING
š ChatGPT Now Wants Your Medical Records
OpenAI opened Health to all US users on Thursday. Connect Apple Health and your hospital records, and ChatGPT will read your labs, meds, and visit notes alongside whatever you ask it.
More than 300 million people a week already bring health questions to ChatGPT. This just hands it your file.
OpenAI says the data wonāt train its models or target ads, and you can disconnect anytime.
The catch privacy experts are flagging: once your records leave a hospital system for a consumer app, HIPAA stops covering them. Read More ā
š INTERESTING READ
The Model that Fits in Your Pocket
While everyone watches trillion-parameter giants, Bonsai 27B went the other way - a full reasoning model squeezed down to one bit per weight, small enough to run on a phone with no internet at all. Worth ten minutes ā
ALSO READ
1. Washington says Kimi cheated. The White House accused Moonshot of training K3 on banned Nvidia chips accessed through Thailand, and of distilling Anthropicās Fable model. Moonshot hasnāt responded. Read More
2. Ask Claude what AI is doing to your job. Anthropic opened its Economic Index as a connector, so you can query real data on AI use across occupations. Read More
3. Meta set an AI optimism ad to a song about human extinction. Nobody appears to have checked the lyrics. Read More
4. Google ships three new Flash models. Gemini 3.6 Flash, 3.5 Flash-Lite, and a Flash Cyber variant, all built for speed and cost rather than the leaderboard. Read More
5. Alexa+ finally reaches your smart home. Amazonās rebuilt assistant can now control the devices itās been sitting next to for a decade. Read More
6. Midjourney bought an astrology app. The image lab acquired Co-Star, which is either a data play or the strangest pivot of the month. Read More
WORTH A TRY AI TOOLS
1. Lucy - Decartās model edits live video as it streams. Swap a character, change the background, drop in an object, all while the cameraās still rolling, in under 40 milliseconds.
Why try it? Live video just became editable. Try it here ā
2. Liso - Highlight any article, newsletter, or thread on the web, right-click, and it reads the thing back to you in a natural voice. Works offline too.
Why try it? Turn your unread tabs into a commute playlist. Try it here ā
Top Tweet of the Week
Someoneās walking into local shops with a Mac Mini and installing private AI in under an hour - no cloud, no monthly bills. The side hustle nobody saw coming.
BEFORE YOU GO
Caption this
The best reply gets featured next week. Just hit reply - we read every one.
See you next week š