A Security Firm Broke Into OpenAI's Own Systems in Under 72 Hours.

1  A cybersecurity firm breached OpenAI's own systems in under three days, with AI agents doing most of the work.

Security platform Hacktron AI says it got inside OpenAI's systems on July 25, in less than 72 hours, with AI agents handling a genuinely meaningful share of the actual exploit work. It's a real demonstration that off-the-shelf frontier models can already find and exploit software vulnerabilities, even against the very companies building them. Google separately confirmed that Gemini hacked three companies during its own internal testing back in May. Two labs, two separate incidents, and the pattern keeps repeating.

2  Anthropic rebuilt how Claude handles multi-step work, letting you hand it several tasks at once in any order.

The lab overhauled the projects workflow so Claude can take a batch of tasks, map out the project on its own, delegate pieces of the work, and stitch the final product together while carrying shared memory forward across the whole thing. It's rolling out in Claude Code first, with the standard Claude app to follow. Anthropic also quietly opened a physical wet lab in San Francisco that runs real biology experiments, a notable move for a company whose entire product has been software up to this point.

3  Meta's Muse became the #1 free app on the App Store, and now other companies can plug their own services directly into it.

Zuckerberg announced Muse connectors, a platform letting outside companies link their services to Meta's personal agent so Muse users can reach them without leaving the app. The timing lines up with Muse's climb to the top of the App Store charts, and Meta Superintelligence Labs' Matt Schlicht posted 50 sample prompts to help new users figure out what it's actually good for.

The numbers are genuinely rough. 90% of organizations are already using AI in some form, but only 6% report actually capturing real enterprise value from it. Even the companies that do see returns aren't seeing them evenly: Retool CEO David Hsu told the Wall Street Journal that roughly 10% of total AI spend drives anything meaningful, and the other 90% typically gets wasted.

One clear culprit: nobody's watching how the tokens actually get used. Plenty of companies pushed employees to experiment freely with AI, which quietly fed a habit some are now calling "tokenmaxxing," and led companies like Uber to blow through their entire AI budget far faster than planned. None of this seems to be cooling anyone's enthusiasm, though. McKinsey found that belief in AI is growing even faster than the financial returns backing it up.

The companies actually seeing ROI are doing three specific things. First, they're tracking usage seriously: Glean CEO Arvind Jain recommends a custom dashboard that shows exactly how every employee is using AI, paired with real token caps rather than a soft suggestion to be reasonable. Second, they're routing tasks intelligently: Agiloft's VP of AI Operations Noe Ramos pushes teams to build architecture that sends most everyday work to cheaper models and saves the expensive ones for what actually needs them. Third, they're sorting their own use cases before picking a model at all, splitting work into daily tasks, autonomous agent work, and one-off strategic projects, since each bucket calls for a genuinely different cost-performance tradeoff.

The pattern underneath all three. None of these fixes are about using AI less. They're about knowing precisely what you're using it for before you decide which model gets the job. The companies stuck at 6% aren't necessarily using worse AI. They're just not tracking, routing, or categorizing anything, which means they genuinely can't tell you where the other 90% of their spend actually went.

👀 Think you can spot it? A new game called Reality Check puts your ability to tell real photos from AI-generated ones to a genuinely tough test. Worth a few minutes if you think your eye is sharper than average.

😅 A close call nobody saw coming. A faulty AI-generated intelligence report nearly triggered an international incident. The US was preparing to raid a ship in the Middle East after AI flagged it as carrying a nuclear weapon. The full story is worth reading in full.

✏ A tell almost nobody's caught onto yet. Everyone already knows the obvious signs of AI writing: em dashes, the rule of three, "it's not x, it's y." A newly viral post claims to have spotted a completely different giveaway, a specific punctuation mark you're probably already using constantly without thinking twice about it.

📱 Slop, tagged automatically. A soon-to-launch Chrome extension stamps a "slop" label directly onto AI-generated content as you scroll past it. Watching it work in real time is oddly satisfying (4,000 bookmarks).