Let’s start today’s tech briefing on a bittersweet note. Over the past few months, ICE agents have been spotted wearing smart glasses equipped with cameras, which raised a lot of privacy alarms. Now, the federal agency is banning employees from wearing such smart glasses on active duty. On a related note, MIT experts say it’s getting increasingly difficult to trace the human inspiration behind AI-generated images. It’s a worrying development for copyright and fair use concepts. And finally, an image search tool exposed a database of millions of humans by dumping it openly on the internet. It seems like we walk a step forward, and take two steps back in the digital age.

Send us feedback at [email protected].

Immigration and Customs Enforcement has told all employees they are not permitted to wear Meta smart glasses while on the job, according to an internal memo reported by The New York Times. The agency warned that the devices "could unintentionally capture, record, or transmit sensitive information, potentially compromising privacy and legal protections." A Department of Homeland Security spokesperson framed the directive as a clarification of a longstanding ban on personally owned body-worn cameras rather than an entirely new policy.

The memo arrives after months of Meta's Ray-Ban smart glasses turning up conspicuously in immigration enforcement settings. Agents from ICE and Customs and Border Protection have been spotted wearing the glasses across at least six states, and in at least one documented case, a CBP agent was recorded using them to capture footage despite the prohibition. The visibility of the glasses during raids and enforcement operations has drawn significant public attention and added to existing concerns about surveillance overreach. Interestingly, ICE is seeking bespoke smart glasses with biometric identification capabilities as soon as next year, suggesting the ban is less about rejecting the concept than about controlling how and when it is deployed.

ClarityCheck, a people-search website that promises users their reverse image searches are "private and secure," left more than 9 million image files openly accessible on the internet, according to research by independent security researcher Jeremiah Fowler. The exposed Amazon S3 bucket contained roughly 450 GB of data stored in folders labeled "faces" and "profiles," reachable by anyone who knew the URL embedded in the company's own publicly available website code. A second misconfiguration separately exposed users' email addresses and phone numbers. The images included photographs of adults, teenagers, and children, sensitive biometric data that, unlike a password, cannot simply be changed if it falls into the wrong hands.

The exposure is especially troubling given what ClarityCheck is designed to do. The service explicitly markets itself as a tool to identify strangers from photos and find their social media profiles in seconds, meaning the people whose faces ended up in the database likely never knew they were there and almost certainly never consented to it. Fowler warns the bucket appeared to have been exposed for months before it was secured, and his initial attempts to alert the company went nowhere. He also flagged that an automated bot could have crawled the database, harvested the face images, and used them to train AI models. ClarityCheck disputes the word "exposed," arguing that finding the URL required specific knowledge not available through ordinary browsing, and has since secured the database.

A new study from MIT's Computer Science and Artificial Intelligence Laboratory offers a striking challenge to one of the central assumptions behind AI copyright lawsuits: that a generated image can be meaningfully traced back to specific training data. Researchers Zheng Dai and David Gifford identified a phenomenon they call attribution decay, finding that as generative models are trained on larger and larger datasets, the influence of any individual image effectively vanishes. Remove a single photograph, remove every image by a given artist, remove every picture of a particular person, and at sufficient scale, the generated output does not change. Their paper published today in Nature Communications is the first to demonstrate this not through approximation, but through exact deletion.

Testing across 24 ensembles and datasets ranging from 256 to more than 160,000 images drawn from collections including CIFAR-10, CelebA, and ArtBench, the team found the same pattern every time: the larger the training set, the smaller the measurable influence of any single piece of data, shrinking along an inverse power law. The legal and creative implications are substantial. Courts and technologists will need to look beyond attribution entirely when assessing whether a model's output constitutes copying. Companies wishing to demonstrate their outputs are not copyright-infringing derivatives must now actively revise their models to guarantee unattributability, framing the finding less as a loophole than as a new technical and ethical standard the industry will need to meet.

Framework Laptop 12 starts at $699 with Intel Core Series 3

Marvel Tokon runs better cracked than officially released

Narwhal tusks have two spirals twisting in opposite directions

Teenage Sex and Death at Camp Miasma skips CGI for matte paintings

AI-assisted fraud is outsmarting standard security tools

Peacock raises its prices for the fourth time in four years

The Duskbloods beta is now available to download

Arc Raiders map conditions shift to match regional peak hours