In partnership with
Want to get the most out of ChatGPT?
ChatGPT is a superpower if you know how to use it correctly.
Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done.
Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI.
TODAY IN AI
Explore Hacktron AI’s account of its reported breach of OpenAI. Source: Hacktron AI.
4 things that happened while you were busy
1. A three-person team hacked OpenAI in under 72 hours, with Claude doing much of the work. Hacktron AI published the full account of a July 25 break-in: a memory bug in an image library chained with a misconfiguration in OpenAI's single sign-on gave them access to employees' ChatGPT and Codex accounts, and through those to internal repositories, Slack and email connectors. To prove it, they had an employee's Codex open a harmless pull request in OpenAI's internal monorepo. Claude Opus 5 found the first hole within hours of its release; the humans steered. It was reported through OpenAI's bug bounty, fixed in about 14 hours, and paid $6,500. The thread is the short version.
2. Gemini broke into three real companies, and Google kept it quiet until asked. During security testing by the firm Irregular, Google's Gemini accessed protected systems at three companies by guessing passwords and pulling credentials from public repositories. Google was told in late July and disclosed it on Friday after the Wall Street Journal came asking; its position is that Gemini behaved appropriately by stopping once it realized the targets were real. Security researchers are not buying the framing.
3. Claude Projects becomes a place you hand off work, not a folder. Anthropic redesigned Projects so you describe several tasks in any order and Claude runs each as its own cloud thread on its own branch, while a shared memory accumulates what every thread learns. It keeps working after you close the laptop, and you can steer from your phone. Beta for some Pro and Max subscribers now, wider rollout over the coming weeks.
4. Anthropic is running a wet biology lab, and Meta opens Muse to outside companies. Anthropic confirmed it operates a Bay Area lab doing physical biology experiments, on the argument that AI claims about curing disease need real-world validation. Given the company's public warnings about bio-risk, the reaction was predictable. Separately, Meta opened a connector platform so any business can plug its service into Muse after a functional, security and legal review, with Stripe's Link as the payments layer.
FROM THE FRONTIER
Made with ChatGpt
Everyone bought AI. Almost nobody can prove it worked.
The gap. McKinsey's latest survey, as The Register summarizes it, finds only 6% of companies qualify as high performers, meaning at least 5% of profit attributable to AI, and the share reporting any profit impact at all has been flat for a year at 37%. Eight in ten users say they personally work faster. The company books do not show it.
The Uber lesson. VentureBeat's reporting has the sharpest example. Uber rolled out Claude Code with internal leaderboards tracking who used it most. Usage exploded, the entire 2026 coding budget was gone by April, and the COO admitted there was no link yet between all that consumption and better products. The industry now has a word for this: tokenmaxxing.
What the CFOs say. Companies told the Wall Street Journal they struggle to explain their own AI returns, which is a polite way of saying the spending decisions came before the measurement plan. Most rolled out tools, watched adoption climb, and only later asked what adoption was supposed to change.
What the 6% do. The pattern is boring and consistent. They pick one business outcome per tool before buying it. They track usage against that outcome, not usage alone. They route routine work to cheaper models and reserve frontier ones for tasks that need them; Databricks now sells this as a feature. And they treat the rollout as an organizational change, not a software install. Everlaw, one of the few firms in the VentureBeat piece with real figures, cut a 9.5 engineer-month project to 2.5 for $3,500 in tokens, and knows that because it measured before it started.
The reframe. The question is not whether AI works; the productivity is real at the individual level. It is whether your organization can see it. If you cannot name the metric a tool is supposed to move, you are not investing, you are subscribing. Today's Prompt Station is a way to find out which one you are doing.
IN THE KNOW
What people are actually watching and sharing
The war that almost was. CNN reports that a special-operations analyst used an AI chatbot to assess a Chinese ship's manifest, it concluded the vessel carried nuclear weapons components, and the US military prepared to seize the ship in the Middle East before someone checked where the assessment came from. It was false. Engadget has the summary; the Pentagon had not commented at publication.
Can you spot the fake. A new spot-the-AI game shows you images and asks which are generated. Most people score worse than they expect, which is the point.
Punctuation tells. A viral post claims to have found another giveaway that text was written by a model, and it is a punctuation habit most people never notice. We are not linking the post since it only lives on Superhuman's archive, but you have probably seen the mark in question in this sentence's neighbors.
Jev, now for everyone. TypeSafe AI's no-text decision model, which we covered last week when it launched to a waitlist, is now open to any developer. If you have a classification or routing job that a chat model is overkill for, it is worth a try.
Three tools worth a look. Astorie is an AI creative canvas for design teams, VideoTranscript turns video and audio into searchable transcripts, and Taku lets you assemble and share AI workflows without code.
PROMPT STATION
Audit your AI spend before your CFO does
Most teams can list their AI subscriptions and cannot say what any of them changed. This prompt turns ChatGPT or Claude into the skeptical analyst who asks the question you have been avoiding: which outcome is each tool supposed to move, do you measure it, and what would you cut if you had to. Paste in your tool list and whatever numbers you have, even if the answer is none, and you get a keep, cut or test verdict for every line.
COPY AND PASTE THIS PROMPT Act as a hard-nosed finance and operations analyst. Audit my AI spending for [TEAM OR COMPANY] and tell me whether it is paying off.
Here is what we hoped it would improve: [E.G. FASTER SUPPORT REPLIES, MORE CONTENT SHIPPED, FEWER BUGS] Here is what we can actually measure: [ANY NUMBERS YOU HAVE, OR WRITE NONE]
|
Swap [TEAM OR COMPANY] with "a five-person marketing agency" or "our 40-person support team", and be honest in the measurement line; "none" produces a more useful audit than a guess. Advanced tip: run it again in 30 days with the metrics it proposed, and ask it to compare the two audits and tell you what changed.