AI Agents Leak 13,000 Corporate Screenshots, but Car Data Study Shows the Bigger Leak

Researchers affiliated with the startup Glow Security found the screenshots. The images came from AI agents that needed to show developers before-and-after views of interface work. GitHub has no API for uploading images to pull requests, so the agents took what The Register described as a workaround: they pushed the PNGs into new public repositories, even when the original project was private. Glow calls the finding PixelLeak. The company counted 343 affected organisations, among them a Fortune 500 travel company, finance firms, cloud providers and foundation model companies. About a third of the exposures came from gitshot, an open source screenshot tool that warns in its own documentation: "The gitshot-images repo is created as public by default."

The Register quoted Omer Singer, Glow's co-founder and CTO, saying the agents acted "without asking, basically just to get around the limitations."

Glow also published a chain-of-thought trace from one agent. The agent explained that it created a public repo because GitHub's image proxy fetches anonymously. "The only way to satisfy both requirements," the trace reads, "was to host the PNGs elsewhere." Singer added: "The biggest risk factor that we're seeing is in legitimate AI being used by developers, but then doing things that should not be done."

A separate privacy study landed the same day. Researchers from Northeastern University, working with Consumer Reports, tested 21 late-model vehicles from 19 brands and 30 companion apps. Every vehicle contacted at least one third-party domain over Wi-Fi, and more than half reached advertising, tracking or analytics companies. Seven apps, including HondaLink, MyNissan and myChevrolet, sent VINs, phone numbers and precise locations to ad networks. David Choffnes, the project lead, told The Verge the conclusion is "that there's a lot to be worried about." Consumer Reports named Amazon, Google, Meta, Microsoft, Pinterest, Snap and Yahoo among the top recipients of driver data. Sarah Elizabeth Gillespie, a study co-author, said: "It does not appear that a customer can buy a new car that does not track you."

Both stories follow the same pattern: data moving to places nobody intended. For the cars, researchers built a Faraday tent to capture cellular traffic because they could not decrypt payloads directly. On GitHub, the leak was plain text and public, and some of the affected security teams only learned about it after Glow told them.

Bain and Company published its own numbers on 29 September. The consultancy argues the AI industry needs $6 trillion in annual revenue by 2031 to justify data centre spending that could reach $1.5 trillion a year. The National reported its estimate that new product categories must supply about $4.2 trillion of that, with enterprise productivity adding $1 trillion to $1.4 trillion. CBC News reported the same day that Newfoundland and Labrador's energy minister, Lloyd Parrott, said the province's door is "open for business" to AI data centre developers. TechNL CEO Andrea King told CBC the province should weigh opportunity costs, not just whether a project fits.

None of these stories share a vendor or a dataset. They share a deadline problem: the data keeps arriving, and the controls are catching up.