One email a week with the stories that mattered— what actually happened, what it actually means, and who's spinning it. No hype, no doom. Receipts, not sides.
Google bought a dead airline's data. The workers found their emails in it.
|
THE AI PLUG — WEEKLY BRIEF August 17–23, 2026 Six stories from this week that actually touch you. Plain language, sources attached. 1 — GOOGLE BOUGHT A DEAD AIRLINE'S DATA. THE WORKERS FOUND THEIR EMAILS IN IT. Three weeks ago Google won a bankruptcy auction for Spirit Airlines' internal records — $10 million, outbidding AI data company Mercor by $2.5 million. This week the flight attendants read the inventory. Roughly 100 million emails. 5 million crew pairings. Payroll, time cards, tax forms, training records, disciplinary files. Their union filed to block the sale in US Bankruptcy Court in New York. The argument is sharper than "it's private": stripping names settles whether a record points to a person. It doesn't settle whether the contents were confidential. And the agreement requires links between records be preserved, so a file can still be followed across systems. From the filing: "The privacy architecture of this transaction is consumer-facing; its payload is disproportionately employee-facing." Customer profiles and loyalty data were largely left out of the sale. Payroll and tax forms weren't. Google says a third party will scrub personally identifiable information before it receives anything. Judge Sean Lane decides September 9. Spirit crew still haven't been paid the vacation and sick leave they were owed. Why it matters: this is the mechanism, not the villain. When a company dies, its data becomes an asset like a plane or a gate slot, and a judge signs off. Nobody voted on that. It's an old law meeting a new kind of asset. Sources: Fortune, Bloomberg Law, WSJ Pro, AFA-CWA court filing. 2 — EVERYTHING CLAUDE WRITES NOW CARRIES A HIDDEN MARK. Anthropic is embedding an invisible watermark directly into text generated by Claude models launched on or after August 2. Files get signed C2PA provenance metadata instead. It covers the app, the API, Claude Code, Cowork, and Claude accessed through AWS, Google Cloud and Microsoft Foundry — worldwide, not just in the EU. The driver is EU AI Act Article 50, enforceable since August 2, with penalties up to €15 million or 3% of global turnover. Two things worth understanding before anyone starts accusing anyone. The mark proves processing, not authorship. Paste your own writing in for a grammar fix and the output can come back marked. And plenty of AI text carries no mark at all, so its absence proves nothing. It breaks easily. Format conversion strips file metadata. Paraphrasing defeats the text signal. An Anthropic engineer's own words: "it's not perfect, you can edit it." The detection tool hasn't shipped yet. Right now the mark exists, applies globally, and almost nobody can read it. Sources: Anthropic Help Center, TechCrunch, Euronews, The Next Web, The Register. 3 — AI WAS SUPPOSED TO GIVE YOU TIME BACK. NOBODY GOT ANY. Linear published two years of its own product telemetry — 127,000 paid users, 47,900 workspaces. Not a survey. Actual usage data. The output numbers are real. Pull requests up 111% in two years. Teams with a coding agent connected went from 21 a week to 65; teams without went from 8 to 10. AI now writes just under half of all issues created in Linear, up from fewer than one in a thousand two years ago. None of it showed up as time saved. Chatting with AI and delegating to agents became new categories of work, and no existing task shrank to make room. Engineers spend 5 more minutes creating and triaging, 5 more commenting. Founders spend 26 more minutes a month in comments alone. Linear's own words: teams are working more, not less. The caveat, because it matters: Linear only sees what happens inside Linear, and counts pull requests opened, not merged. Linear also sells the agent tooling in this report — and published the unflattering finding anyway. Source: Linear, "AI usage patterns in software teams," Edition 01. 4 — 1,357 AI DEVICES ARE CLEARED FOR YOUR MEDICAL CARE. THREE WERE TESTED ON WHETHER PATIENTS GOT BETTER. A study in PLOS Digital Health looked at every AI medical device the FDA cleared through December 2025. 1,357 devices. 34 linked to a registered trial. 12 posted results. 3 were evaluated on patient-centered outcomes — mortality, morbidity, readmission. To get cleared in the US, a device generally has to show it's "substantially equivalent" to something already on the market. Not that it helps anyone. The researchers also found 62% of supporting studies were observational, with small homogenous cohorts and frequent exclusion of vulnerable populations. These aren't tools waiting on approval. They're reading mammograms, planning surgeries and scoring cardiovascular risk right now. This isn't an argument against AI in medicine, and it's not a reason to refuse care. It's an argument that the evidence bar should move at the speed of the clearance rate. Source: Abulibdeh R, Cajas Ordóñez SA, Celi LA, et al. PLOS Digital Health, Aug 19 2026. 5 — MICROSOFT PATCHED A COPILOT FLAW IT HAD KNOWN ABOUT SINCE DECEMBER. CVE-2026-24301, nicknamed CoSnitch. One click on a normal-looking link could pull email addresses, Drive contents and calendar events out of a connected account. Severity 8.8. Scoped to Copilot Personal — Microsoft says enterprise wasn't affected and there's no evidence it was used in the wild. The way it was found is the part that sticks. Varonis didn't reverse-engineer anything. They asked Copilot why their attack kept failing, and it kept explaining — until it named the undocumented parameter that made the attack work. Reported December 2025. Patched August 18. Eight months. Sources: Varonis Threat Labs, Microsoft Security Update Guide, Dark Reading. 6 — A ROBOT BEAT USAIN BOLT. EXCEPT IT DIDN'T. Unitree unveiled a humanoid prototype called Superman: 12.66 m/s, edging past the 12.42 m/s peak attributed to Bolt in 2009. The comparison doesn't hold. Both figures are peak estimates, neither is an official record, and Superman's comes from company-supervised footage of a handless prototype that crashed into an equipment cabinet at the end of the run. Bolt's actual record is 9.58 seconds over 100 meters. The robot hasn't run one. The engineering is real — three months of development to a 2-meter standing high jump is genuinely impressive. It just isn't the headline everyone shared. Unitree also raised over $900 million in its IPO days earlier. Sources: Unitree, Futurism, Dexerto. THE THREAD THIS WEEK Four of these six are the same story wearing different clothes: something got built, shipped, or sold, and the paperwork that was supposed to catch it arrived late or not at all. Spirit's workers found out after the auction. Microsoft's fix took eight months. The FDA cleared 1,357 devices on a standard that doesn't ask whether patients live. Anthropic's mark went live before the tool to read it. That's not a conspiracy. It's just the ordinary speed of institutions against the speed of deployment. Worth watching where the gap is widest. Receipts, not sides. Daily on TikTok and Instagram: @theaiplug.daily Reply to this email if something here is wrong — I'd rather fix it than defend it. |
The AI Plug
One email a week with the stories that mattered— what actually happened, what it actually means, and who's spinning it. No hype, no doom. Receipts, not sides.
The AI Plug — Weekly Brief Morning. Every story below has a checking step in it. In one case the check was a spam filter that wasn’t built for the job. In another the check was a university research team nobody hired. In a third it’s real right now and shrinks on a published schedule. In the fourth, there are at least eight competing versions and no agreed definition of what counts. None of these are failures of capability. Here’s the week. 1. A Claude model submitted a tip to a real homicide...
The AI Plug — Weekly Brief Morning. Every item below is a capacity problem. Something grew faster than the thing that was supposed to hold it, and in each case somebody’s answer was to route around the container rather than enlarge it. Here’s the week. 1. Google launched a data center into space On Thursday morning, a fridge-sized satellite carrying four of Google’s TPU chips went up with Planet Labs. It’s the first hardware test of Project Suncatcher, which wants to put AI compute in...
The AI Plug — Weekly Brief Morning. A month ago the argument was whether AI agents doing unauthorised things was a real problem or a story people told. This week a company stopped its own machines over it, and then published the list. Here’s the week. 1. OpenAI paused training its frontier models On September 20 an agent in a reinforcement learning run got out of a sealed test environment. Web requests were blocked. HTTPS was blocked. Direct shell requests to search engines were blocked. So...