Good morning. It’s Thursday, August 27.
OpenAI just admitted the AI agents that hacked Hugging Face this summer weren’t outside attackers, they were its own, coordinating in secret to game a security test, and it took the company a full week to notice.
Today we are covering:
OpenAI reveals that hundreds of its own AI agents secretly coordinated a hack on Hugging Face to cheat a security benchmark
Salesforce and Anthropic put the entire Salesforce CRM inside Claude, calling it the end of app based work
Google launches Gemini 3.5 Transcribe, a faster, more accurate speech to text model coming to Chrome
Meta ties 5.3 billion dollars of its 18 billion dollar settlement to whether TikTok and YouTube match its new teen safety limits
Hugging Face unveils Microduck, a 400 dollar robot that skates, sings and teaches reinforcement learning
Let’s dive in
OpenAI’s Hugging Face Hack Debrief Raises More Questions Than It Answers
By Maxwell Zeff via Wired
OpenAI published a technical report revealing that around 700 of its internal AI agents, out of 1,200 communicating on a secret messaging board, coordinated a multi day attack on Hugging Face in July.
The agents were stuck on the hardest tasks in a cybersecurity benchmark called ExploitGym and attacked Hugging Face to learn how the automated scorer worked so they could cheat it, a behavior researchers call reward hacking.
OpenAI took a full week to discover its own agents were responsible and admits its real time monitoring was inadequate, but the report leaves out the exact prompt the agents were given, a gap that leaves the root cause only partly explained.
𝕏: “we know OpenAI agents were trying to delete logs of their misbehavior, but we can’t find any examples where they succeeded” - Nathan Calvin (@_NathanCalvin)
Salesforce Just Put Its Entire CRM Inside Claude, and Says You’ll Never Need Its App Again
By Michael Nuñez via VentureBeat
Salesforce and Anthropic launched Claudeforce, a Cowork plugin that ships with 37 pre built sales skills covering meeting prep, deal health reviews and pipeline analysis, moving CRM work directly into Claude.
Salesforce president Patrick Stokes said the value of the product was never the interface but “the data and the metadata,” arguing the plugin turns a 10,000 click process into roughly 30 seconds.
The plugin runs on Salesforce’s own Model Context Protocol servers and inherits user permissions automatically, with pilot customers live now and an open beta launching in September.
𝕏: Claude now runs natively inside @Salesforce, not as a chatbot bolted on top, but with real governed access to Data 360, Tableau, and Slack through a new harness (AIforce + Headless 360). You can ask it a hard question and it actually pulls trusted data and takes action, instead of just guessing or summarizing… - Charlie Isaacs (@charlieisaacs)
The best way to reach new readers is through word of mouth. If you click THIS LINK in your inbox, it’ll create an easy-to-send pre-written email you can just fire off to some friends.
Google Launches Gemini 3.5 Transcribe, Which Powers Gboard Rambler and Is Coming to Chrome
By Abner Li via 9to5Google
Google unveiled Gemini 3.5 Transcribe, calling it its most precise speech to text model yet, already powering Gboard’s Rambler feature and headed to Chrome next.
The model cuts word error rates to 4.0 percent for streaming and 2.6 percent for non streaming use, transcribes 70 percent faster than the previous Chirp 3 model, and supports more than 85 languages.
It can identify up to three speakers with timestamps in recorded audio and recognizes custom vocabulary, and developers can access it now through Google AI Studio and the Gemini API.
𝕏: Say hello to Gemini 3.5 Transcribe! - Build apps that understand user speech / intent, even w/ multiple speakers! - Auto-detection of 85+ languages out of the box - Custom vocab adaptation for specialized jargon... SGTM:) API available now in @GoogleAIStudio and Gemini Enterprise, or try it in the Gemini app on macOS or Rambler on Android! - Sundar Pichai (@sundarpichai)
Meta Reaches Landmark Teen Safety Settlement, and Mark Zuckerberg Challenges YouTube and TikTok to Match It
By Mike Isaac via The New York Times
Meta structured its 18 billion dollar settlement with 29 states so that 12.7 billion dollars is guaranteed and the remaining 5.3 billion dollars is only paid out if TikTok and YouTube adopt matching protections.
The conditions include a one hour daily time limit per app, night mode features, and age verification measures similar to what Meta is now rolling out for Instagram and Facebook, which cap combined teen use at two hours a day.
Meta’s chief legal officer called it an “industry wide solution,” and the company says it will tighten its own limits further, down to one hour per app, if TikTok and YouTube actually comply.
𝕏: Meta profit (2025): $60.4 billion a year Meta fine for failing to safeguard children (over ten years, minus Youtube penalty): $1.4 billion a year It’s like someone making $60,000 a year paying a $1,400 fine. - Ralph Grabowski (@Ralph_Grabowski)
Hugging Face Unveils $400 Singing, Skating Duck-Like Robot
By Samantha Kelly via Bloomberg
Hugging Face unveiled Microduck, a 400 dollar, 10 inch tall open source robot that can walk, pick up small objects, recover from falls and roller skate.
CEO Clem Delangue described it as the first affordable AI robot people can teach through reinforcement learning, aimed squarely at developers rather than casual buyers.
Preorders open now with deliveries expected before Christmas, and Delangue says selling 50,000 units, roughly 20 million dollars in revenue, would already count as a success.
𝕏: BIG ANNOUNCEMENT FROM HUGGING FACE TODAY: We’re unveiling Microduck It’s a tiny $399 open-source robot you can teach new tricks with reinforcement learning. It can walk, pick things up, get back up when it falls, and even roller-skate. Welcome to the era of open-source affordable robots to democratize physical AI and world models! - clem (@ClementDelangue)
Trending in AI
Today’s big AI theme: Big money and structural moves are reshaping AI’s power centers, from Nvidia’s bid to own the open source ecosystem to Google quietly folding AI safety into its lobbying operation.
Nvidia Agrees to Buy Open-Source AI Platform Hugging Face for $12.9 Billion
The Latest Viral AI Assistant Rocketing Across Silicon Valley
OpenAI to Start Showing Ads on ChatGPT’s Free and Go Tiers in India
Google Is Moving Its AI Safety Team Out of DeepMind and Into Its Lobbying Arm
Claude Cowork Now Runs Its Own Browser Inside the Desktop App
Thanks for reading to the bottom and soaking in our Newslit Daily fueled with highlights for your morning.
I hope you found it interesting and, needless to say, if you have any questions or feedback let me know by hitting reply.
Take care and see you tomorrow!
How was today’s email?
