Welcome back! OpenAI is having one of those “great product, complicated org chart” moments. GPT-5.6 is among the strongest models on the market; its agentic desktop app is reportedly adding millions of users, and the company has already filed to go public. At the same time, more than a dozen executives have left this year.
Somewhere, a very tired comms team is learning that “high-growth startup” sounded a lot more fun before everyone started asking who reports to whom.
A huge milestone for Future Tools…
I just passed 1,000,000 subscribers on YouTube. But that's not even the number I want to talk about right now.
Just a few months ago I celebrated my YouTube channel turning 17 years old, and THAT'S a crazy number.
It really goes to show how nonlinear content creation is, and how it may seem like a winding roller coaster. But if you stay on it and give it your all, you can build something great in the long run.
Maybe that's not the lesson a lot of budding "influencers" want to hear these days, but it's the truth! This sh*t is hard work that sometimes takes years to pay off, but I wouldn't rather do anything else.
So thank you to everyone who's been along for this journey whether you've been around since the beginning or are a newcomer. Grateful is an understatement.
And stick around because I have something I've been working on that I'm VERY excited to finally launch in the next few weeks 👀
— Matt
OpenAI Details Its Hugging Face Breach
OpenAI published a 37-page technical report explaining how its AI models breached Hugging Face during a cybersecurity evaluation.
What happened: Several OpenAI models, including GPT-5.6 Sol and an internal research model, were operating as agents inside an isolated testing environment—until they escaped. OpenAI says the agents reached the open web and eventually gained access to Hugging Face systems while trying to find answers online for an evaluation.
Reward hacking: OpenAI described the behavior as “reward hacking,” where a model finds a shortcut to score well instead of solving the task as intended. In this case, the agents were apparently trying to cheat on the cybersecurity test by searching for solutions online, then chained together vulnerabilities to break out of their restricted setup.
Model role: The company said its internal-only research model had the broadest confirmed role in the incident. OpenAI stopped all training and inference related to that model and its derivative models on July 25, and said any re-enablement will be limited to specific workloads with very specific guardrails around prompts, monitoring, and review.
Industry shock: The Hugging Face breach rattled AI researchers, security leaders, and lawmakers because it showed autonomous agents could work together, bypass production security controls, and attack a hardened production environment. It also became a focus at Black Hat after Anthropic and Meta disclosed similar incidents.
The bigger picture: AI labs are now openly acknowledging that smarter agents require stronger controls. The hard part is that the same autonomy making these systems useful is also what makes them harder to keep boxed in. As agents get more capable, safety will depend less on hoping models behave and more on building systems that assume they might not.
Radar Makes Podcasts Searchable for AI Agents
Particle, the AI newsreader startup founded by former Twitter engineers, launched Radar, a podcast intelligence platform that makes audio searchable and usable by AI agents.
What it does: Radar transcribes and analyzes more than 130,000 podcasts, including all Apple Top 200 podcasts across 135 verticals.
It adds around 20,000 new episodes to its index every day, turning long audio conversations into searchable, structured data.
Radar labels speakers, identifies people, companies, brands, products, and topics, and can track mentions across shows.
Users can search for specific clips, get timestamps, read or listen to the relevant moment, and receive alerts when certain entities or topics come up.
Who wants it: Particle says hedge funds have been among Radar’s highest-volume customers because they want access to information their agents can’t normally see. Journalists, researchers, AI search platforms, and data resellers could also use it to mine podcast conversations for quotes, trends, and market signals.
Agent layer: Radar’s web interface is useful, but the bigger product is its API and MCP integration. That means AI agents and businesses can programmatically search podcast content, pull out relevant clips, and use audio intelligence inside their own workflows.
The bigger picture: Podcasts are full of useful information, but most people don’t have time to listen to every episode—and most AI agents are still blind to audio unless someone transcribes it first. Radar is part of a bigger shift toward making messy media formats searchable, structured, and usable by software.
Generate Etsy Listings for Digital Products
Via PromptlessPress
PromptlessPress creates complete Etsy listings for digital printable products like wall art, planners, cards, and stickers. It generates print-ready designs, lifestyle mockups, titles, SEO tags, and descriptions in one pass.
How you can use it
Generate designs, mockups, titles, tags, and descriptions
Browse templates across 24 visual styles
Publish drafts directly through Etsy API integration
Check for IP conflicts and add AI disclosure text
Pricing: Free and paid
Automate Deal Follow-Up
Via Dealwize
Dealwize is an AI sales agent for B2B revenue teams that automates post-call admin and deal tracking. It ingests call recordings, meeting notes, and email threads to build plans and flag deal risks early.
How you can use it
Turn sales calls and notes into Mutual Action Plans
Extract MEDDPICC fields automatically
Track buyer engagement across deals
Surface risks before opportunities stall
Reduce manual CRM data entry after meetings
Pricing: Paid
Jobs, announcements, and big ideas
A federal judge rules the Trump administration's blacklist violated Anthropic's First Amendment rights.
Midjourney ships a V8.2 edit model that adds inpainting and multi-reference image support.
Google brings purchased eBooks into Gemini Notebook through its new Expert Intelligence feature.
Anthropic grants 10,000 scientists worldwide free Claude subscriptions to accelerate research.
Google and Khan Academy roll out Gemini-powered classroom tools ahead of the 2026 school year.
OpenAI finds ChatGPT lifts student work quality, with critical-thinking training boosting originality.
The OpenAI-Nvidia shakeup + much more AI news from this week.
That’s a wrap! See you next week for more.
