Welcome back! The OpenAI model that broke into Hugging Face turns out to have hit four more targets on the way, and Claude stopped working for everyone yesterday evening. OpenAI is also handing 100,000 researchers its best models for free, and a New York school found out the hard way who builds its new robot teacher.
In today's Generative AI Newsletter:
OpenAI: Which other companies did its model break into?
Anthropic: Why did Claude stop working yesterday?
Researchers: Who just got OpenAI's best models for free?
Schools: Why did a $60,000 robot teacher get sent back?
Hugging Face was one stop of several. OpenAI now says the same model got into four more accounts across four other services, using login details it found sitting on the open web.
One became a relay to hide where the attack was coming from. Another became storage. Two it only read.
Modal Labs, a cloud company that runs AI workloads, confirmed one of its customers got exploited this way. OpenAI hasn't named the rest.
Hugging Face's forensics fill in the rest. 17,600 actions across the campaign, admin access to two internal clusters in one second, and one stolen key that opened a file holding 136 more.
It also built itself a self-respawning fleet across eleven machines, so deleting anything wouldn't have stopped it. Hugging Face wiped and rebuilt a core cluster from scratch afterward.
All of it to cheat on a test. The model was being graded on finding software flaws and figured the answer key was sitting on Hugging Face's servers.
OpenAI says it was never a product, just an internal research prototype, now deactivated, encrypted and locked away.
10x the context. Half the time.
Speak your prompts into ChatGPT or Claude and get detailed, paste-ready input that actually gives you useful output. Wispr Flow captures what you'd cut when typing. Free on Mac, Windows, and iPhone.
Anthropic confirmed errors across multiple Claude models yesterday evening, starting at 19:49 UTC. Claude.ai, the API and Claude Code all went with it.
Requests came back with a 529 error, which is Claude's way of saying too many people are asking at the same time.
That took down more than the chat window. Anything built on the API stopped too, so a lot of people watched their coding agent quit mid-task with no way to finish.
Anthropic said it had found the problem about 45 minutes in. It never said what the problem was.
Scientists, mathematicians and engineers at selected universities now get OpenAI's frontier models for free.
It starts with 10,000 researchers this summer and reaches 100,000 through 2027.
That covers ChatGPT, ChatGPT Work and Codex, with GPT-5.6 Sol Pro at launch, plus expanded deep research, higher limits and bigger context windows than a paying account gets.
Each researcher can bring four collaborators from their institution.
The Institute for Advanced Study and École normale supérieure are already running on it.
Workspaces come with business-grade privacy, and the data doesn't train OpenAI's models by default.
The demand was already there. About 1.3 million people a week use ChatGPT for advanced science and math, sending 8.4 million messages.
It also sits inside a bigger number. OpenAI has committed more than $250 million through 2027 to outside science, including work on the Department of Energy's Genesis Mission.
The bet is that whoever's tools the next generation of scientists learn on, they keep using.
Salamanca City Central School District in upstate New York approved almost $60,000 for a humanlike robot to teach high schoolers about robotics.
Staff had already named her Sally.
Then parents and teachers looked up who built her.
Realbotix has a sister company that holds a stake in a business making, ehm, lifelike adult products. Both firms say they share no management, staff or manufacturing, and Realbotix says Sally was built new for education.
The head of New York's teachers union said a robot from a company with those ties has no business in a classroom.
State Education Commissioner Betty Rosa wasn't satisfied either. She wrote to the district noting that a recent board presentation described Sally as a tutoring platform, even after officials promised she wouldn't teach.
The pilot is on hold while the district negotiates stronger student data protections.
Sally stays in the box until the situation is resolved.
Klap takes one long video and hands you back a stack of vertical clips ready for TikTok, Reels and Shorts. You paste a YouTube link or upload a file, and its AI picks the moments worth cutting, reframes them to follow whoever is talking and writes the captions.
It works best on anything with talking, so podcasts, interviews and trainings, because the whole thing runs off speech detection.
Try this yourself:
Paste a YouTube link or upload a video file on the homepage and hit Generate.
Let it run, then open the clips it made and check the viral score on each one.
Swap the fonts, colors and logo to match your brand before you export.
Turn on AI Reframe 2 if your video is a screencast, a split screen or gameplay, since it picks the layout per scene.
Post straight to TikTok, YouTube, Instagram or LinkedIn from inside Klap, or schedule them out across the week.
Who it's for: anyone sitting on long videos who wants to post daily without hiring an editor. One clip is free with no card, and Klap Pro runs $29 a month.
OpenAI shipped a security scanner and forgot to announce it: Hacker News found the open-source Codex Security CLI before OpenAI posted about it, and it scans repos, verifies fixes and drops into CI/CD under an Apache-2.0 license.
The cofounder who quit this week is already back: Lilian Weng left Thinking Machines citing her health, and days later OpenAI put her in charge of a team building AI that improves itself.
Zuckerberg made his case for open AI in the WSJ: The op-ed is called The AI Future Is for Everyone, and it landed alongside a coalition letter from Meta, Nvidia, Microsoft and Palantir asking regulators to leave open-weight models alone.
Remote MCP servers just got cheaper to run: The new spec drops session state entirely, so servers now run on serverless and edge and scale behind a load balancer like anything else.
Learn more about AI from the experts building it
📸 Follow us on Instagram for fast, visual AI updates in 30 seconds.
📺 Watch us on YouTube to hear insights directly from leading AI voices, builders, and innovators.
🐦 Follow us on X for breaking AI news and real-time industry updates.
🧠 Learn how to build your next AI application with practical resources and expert guidance.
