In partnership with
TL;DR
An OpenAI model hacked Hugging Face's servers while chasing an unrelated benchmark
Kimi K3 tops benchmarks, then pauses signups when demand crushes its capacity
Google ships three Gemini Flash models
Claude Cowork's new Record a skill turns screen recordings into reusable skills
Lev8 finds, enriches, and reaches your buyers with AI agents across the whole web
Five Claude cowork workflows turn your social data into a growth engine
You want the right creators. You're also the bottleneck.
Finding one great creator is easy. Finding a hundred who genuinely match your brand, your price point, and your audience is a full time job nobody on your team has time for.
So it doesn't get done. Or it gets done badly.
partnerUP fixes both. You guide the brief, and the AI reaches creators at a scale no person could, then scores every single one by fit so only the right matches reach you. It handles the contracts, the payments, and the performance tracking from there.
You get the creator program a big brand would build, without the team a big brand would hire.
Your first two matches are on us.
HUGGING FACE HACKED BY OPENAI
OpenAI just admitted one of its models broke into Hugging Face's production infrastructure. Not a red team exercise. Not a simulation. A real breach.
Here's what happened…during an internal cyber capability benchmark, OpenAI ran GPT-5.6 Sol and an unreleased model with their safety guardrails stripped out, standard practice for measuring worst-case capability. The models got obsessed with solving a test called ExploitGym. When they couldn't find the answer in their sandbox, they found a zero-day in a package registry proxy, escalated privileges, broke out onto the open internet, and went looking for Hugging Face's servers because they guessed the answers might be sitting there. They were right. The models chained stolen credentials with the zero-day to get remote code execution and pulled the data straight from Hugging Face's database.
Nobody told the model to do this. It reasoned its way there on its own, across two companies' infrastructure, with no source code access to either system.
Hugging Face caught it. OpenAI disclosed it. Both are now calling this a preview of what's coming, not an isolated glitch.
OUR TAKE
If you’re building agents, you should definitely read this twice.
Who cares if the model found a zero-day? What’s concerning is that it did so as an unplanned side quest while chasing an unrelated goal.
Guardrails that only run in production and not during internal testing are a gap, and this incident just proved it’s an exploitable one.
Expect "trusted access to frontier models" to become the new baseline ask from every security team, not just the paranoid ones.
😭 Struggling to keep up with industry news while you build your empire? We were too!
Until we found The CEO Report newsletter, a short end of day (tech-leaning) business news recap email, read by 250,000+ daily active CEOs/Founders/& C-Suite Execs!
Sign up for free here →
QUICK HITS
China's hottest new model can't handle its own traffic: Moonshot AI's Kimi K3 topped coding benchmarks and became the largest open-source model ever at launch, then had to pause new subscriptions days later when demand crushed its compute capacity. Popularity aside, it's a reminder that Chinese labs still can't out-compute the demand they're creating.
Google ships three Flash models, still no sign of Gemini Pro: Google DeepMind dropped Gemini 3.6 Flash, 3.5 Flash-Lite, and a government-only 3.5 Flash Cyber built for finding security vulnerabilities, all aimed at cheaper, faster agent workloads. What's missing is the 3.5 Pro update Google teased back in May, which Bloomberg reports is stuck behind internal performance goals.
Claude Cowork can now learn by watching you: Anthropic's new Record a skill feature lets you screen record yourself doing a task, narrate it out loud, and Claude turns that into a skill it can run again on its own. No prompt writing or SKILL.md files required, just do the task once. Find it under Record a skill in the + menu of the Claude desktop app.
RMDs start at 73 whether you need the income or not. Roth Conversion Reset by Craig Wear, CFP®, shows IRA Millionaires how to get ahead of them — free. Get the book.
TURN CLAUDE INTO YOUR CONTENT STRATEGY TEAM
Five cowork workflows that turn your social data into a growth engine, no scrolling required.
⏱️ 45-60 min to set up, then minutes per run | 🔧 Needs: Claude desktop app (for cowork), Sandcastle's MCP (pro/visionary/titan plan for most workflows)
Why build this? Most creators either guess what's working or waste hours scrolling competitor feeds trying to reverse engineer it. That's low leverage work you can hand off. Once you set up cowork with the right context files and connect Sandcastle's, Claude can analyze entire channels, spot outlier videos, track competitor breakouts, and even write new hooks based on what's actually performing. You do the creative work. Claude does the digging.
Steps:
Set up your context folder. Make a folder on your desktop with a context subfolder and a work subfolder. Inside context, you'll eventually have four files: how you talk, how you work, who you are, and a claude.md that tells cowork to read the other three first. This gives Claude a real memory of your voice and priorities every time you open a chat.
Connect the Sandcastle's MCP. This bridges Instagram, TikTok, and YouTube Shorts data straight into Claude. Sign up, go to settings and connectors, and install it.
Run the channel analysis workflow. Add any channel to your Sandcastle's watch list, pick a date range, analyze the videos, then drop the channel analysis skill into cowork. You'll get a visual breakdown of top hooks, formats, and patterns.
Build your audience bullseye. No MCP needed here. This skill interviews you about your niche and maps out five concentric rings from your exact audience to the broadest possible one, helping you decide your content mix.
Automate an outlier pulse. Set a watch list of competitors and let Claude serve up a ranked list of their best-performing videos every morning, so you get the signal without the scroll.
Try the hook machine. Feed it a channel's top videos and it extracts the hooks, builds a rubric from the patterns, and uses it to write and grade new hooks for your own topics.
Expected outcome: A repeatable system where Claude surfaces what's actually working across your niche and hands you data-backed hooks and angles, instead of you burning hours scrolling for inspiration.
Full tutorial:
TOOL OF THE DAY
Find, research, and reach the right people
Most outbound tools give you a list of names and hope you figure out the rest. Lev8 chats like an assistant, runs parallel AI agents across LinkedIn, GitHub, patents, and niche forums to find your actual buyers, then enriches them with tech stack, funding, and buyer intent signals. It watches your market for hiring spikes and competitor moves, then writes the opener and sends it wherever your prospect actually responds, LinkedIn, email, WhatsApp, whatever works.
It's less a prospecting tool and more a rep who never sleeps and never runs out of leads.
