Big Tech’s Seventeen Most Interesting New Patents This Week (8/28/2026)

A weekly read of what the world’s biggest technology companies just told the U.S. patent office.

Another heavy week: we wrote up 285 filings, with Samsung out front at 55, then Qualcomm at 36, Sony at 33, and Microsoft at 30. The full run lives under this week’s Big Tech patents; these seventeen stood out.

1. IBM Patents Technology That Links Brain Signals Directly to Search and Retrieval

You remember reading that email on a rainy afternoon in a coffee shop. Your inbox does not. IBM has filed for a system that records your brain activity alongside what is happening around you, then lets you search the bundle later.

A neural interface, a wearable or an implant, reads electrical signals from your brain while your device’s sensors log the outside half: where you are, what you are doing, what the room looks like. Each moment gets saved as one record with an inner half and an outer half.

Later you type an ordinary search. If the query touches a captured moment, the system returns both the event and your brain state at that instant. Memory becomes a filing cabinet with your reactions filed alongside the facts.

The catch is everything around the edges. Hardware that reads brain signals cleanly all day does not exist outside a lab, and an archive of your inner responses is the most sensitive database anyone has ever proposed building. Who gets to query it is a harder problem than building it.

2. Meta Patents Smart Glasses That Read Your Silent Mouthed Words

Talking to a voice assistant in a library is awkward. In a loud crowd it is useless. Meta has filed for glasses that understand words you mouth without making a sound.

When you mouth a word, your face and jaw muscles still move in recognizable patterns even though no air moves. Sensors pressed against your head pick up those contractions and tiny vibrations, and software translates them back into the words you intended. Even a partial lip movement can carry a command.

For a device that sits on your face all day, this is the missing input method: invisible to the person next to you, usable anywhere speech is not.

Last week Meta filed a chip-level mute switch that no software can override; this week it files a sensor that hears words you never said out loud. The company is drawing both ends of the privacy map at once, and the hard part here is mundane: telling a silent command apart from chewing, and convincing anyone the microphone-that-is-not-a-microphone stays off.

3. Nvidia Patents a System for Steering Autonomous Vehicles via VR Headset

Every self-driving car eventually meets a construction zone it cannot read, and today the answer is a stopped car and an apology. Nvidia has filed for handing the wheel to a human miles away wearing a VR headset.

The operator sees a live 3D view built from the car’s cameras and sensors and looks around by moving their head. The system streams only the slice of sensor data that matches where the operator is facing, which keeps the connection from drowning. Steering and speed inputs go back to the car in real time, and control returns to the software once the mess is behind it.

For anyone running a robotaxi fleet, this is the difference between a rescue driver in a headset and a rescue truck on the highway.

The obvious weak point is the link itself: latency or a dropped connection at the exact moment a human took over is the nightmare scenario, and the filing lives or dies on how it degrades. One more detail worth the read: the first name on the inventor list is Jen-Hsun Huang, Nvidia’s CEO.

4. Microsoft Patents a Computing System That Uses Light From Transparent Displays as Its Processor

Every computer you have ever used moves electrons through chips. Microsoft has filed for a machine where the light from a stack of transparent screens is the processor.

Picture a rack of see-through displays, each driven by its own small controller, each glowing with a precise pattern. A sensor watches all of them at once and reads their combined glow as a single signal. That combined reading is the answer to the calculation. The screens do not show the result of the computation; their light is the computation.

The appeal is massive parallelism without shuttling data through wires, which is where much of modern computing’s power bill goes.

This is research-stage thinking, not a product plan: a single named inventor, a continuation filing, and optics that would need to hold alignment in the real world. But when a company that runs planet-scale data centers files on ways to compute without wires, the direction says something.

5. Sony Patents Sunglasses That Project Floating Images Into Your Field of Vision

Every AR headset so far has looked like a headset. Sony has filed for sunglasses that paint floating images onto the lens using invisible ultraviolet light.

UV emitters hidden in the frame shine into the curved lens, where microscopic phosphor dots, the same family of material that made old TV screens glow, convert the invisible light into visible colors. The image forms in the lens itself, so there is no projector box clipped to your temple.

That is a genuinely different answer to AR’s oldest problem: making the display hardware disappear. Thin, light, all-day wearable, with a wide field of view, and the sunglass tint helps the contrast for free.

The filing comes from Sony Interactive Entertainment, the PlayStation division, and continues an application from 2022, so someone has believed in this for a while. What it has to prove is brightness: phosphor dots against a sunny street is a hard fight, and paper does not settle it.

6. IBM Patents an Automated System for Finding Gaps in AI Safety Filters

Every trick that talks a chatbot past its safety rules was found by a person, one clever attempt at a time. IBM has filed for a machine that runs the whole jailbreaking process automatically.

The system collects known bypass tricks, groups them by style, then mixes and matches the styles to generate thousands of fresh test prompts. It filters the duplicates, fires the rest at the target AI, and scores which ones actually got through. The output is a ranked list of the prompts that made the model misbehave.

The point is defensive: find the holes in your guardrails before your users do, at a scale no human red team can match.

The uncomfortable symmetry is that this is also a working description of an attack machine, and the patent cannot control who runs one. Safety testing at industrial scale means jailbreak generation at industrial scale; the only question is whose payroll the machine is on.

7. Nvidia Patents a System That Lets AI Models Select Their Own Training Data

Training a large AI model starts with a guess: how much code, how much news, how many scientific papers. Guess wrong and you burn months of compute. Nvidia has filed for a system where a model works out its own training diet.

The raw text pile gets sorted into clusters by topic and type. A small helper model then runs rounds of trial-and-error experiments, each round building on the last, to learn which blend of clusters produces the best results. The output is a recipe, a weighted formula for the mix, and the full-size model trains on that.

Data mixing has been expensive human guesswork at exactly the moment it matters most. Automating it means faster iteration on every model that follows.

The trade is that the recipe inherits the helper model’s blind spots, tuned toward whatever the trials measured. A model that picks its own diet eats what scores well, and what scores well is not always what nourishes.

8. Microsoft Patent Would Auto-Summon Absent Colleagues When Meetings Require Their Input

Every meeting invite includes a few people listed just in case, and they never join. Microsoft has filed for an AI agent that attends in your place and decides when to drag you in.

The agent sits in the meeting as what the filing calls a ghost presence, listening for moments that involve you. It has been trained on your expertise, history, and habits, so it can judge two things: whether a question is really yours, and whether the room can resolve it without you. Only when both answers point at you does it reach out, by message, call, video link, or, the patent notes, pager.

Done right, this is the end of attending a full hour for your four minutes. You go dark on the meeting without going dark on the decision.

Done wrong, it is a machine that interrupts you with other people’s meetings, calibrated by a model’s opinion of what you know. The threshold is the whole product, and the training data is a profile of you that your employer maintains.

9. Apple Patents a Way to Pay With a Friend’s Card While Keeping Your Own Loyalty Perks

Your partner’s card has the cashback. Your account has the discount code. Today you pick one or hand a physical card across the table. Apple has filed for paying with someone else’s card from your own phone, benefits intact on both sides.

At checkout, your iPhone shows their card as an option. Selecting it sends the order details to their device, which generates a one-time authorization token, a narrow cryptographic proof derived from the card, and sends it back. The card number itself never leaves their phone, and they approve the specific purchase on their end.

Their card benefits ride the token; your merchant perks stay attached to your account. It is a signed check for one purchase instead of a borrowed wallet.

The elegant part is on-device. The hard part is everything else: banks and card networks have fraud rules built on the assumption that the cardholder is the payer, and second-party payments break that assumption on purpose.

10. Microsoft Patent Targets Hackers Who Trick AI Into Revealing Forbidden Information

A security camera does not understand intruders; it compares shadows against a library of silhouettes. Microsoft has filed for the same idea applied to AI chatbots: screen every prompt against a library of known attacks before the model ever sees it.

Each incoming message is converted into an embedding, a numerical fingerprint of its meaning and structure. That fingerprint is scored for similarity against a database of fingerprints from past attack prompts, the phrasings that tricked models into leaking instructions or dropping their rules. Too close a match and the message is stopped at the door.

The payoff is a filter that learns from every attack it has already survived, closer to antivirus software than to a content filter.

Antivirus history also tells you the weakness: signatures catch variants of what is known, and the attacker who invents a genuinely new phrasing walks past. It raises the floor. Nobody should mistake it for a ceiling.

11. Sony Patent Generates Personalized Store Ads From Your In-Store Movement

You pause at the coffee display, double back to the snack aisle, linger near the protein bars. To the store, that has always been invisible. Sony has filed for a system that turns your movement through a store into ads generated for you on the spot.

One part watches behavior in a defined space: where people go, what they stop near, how long they stay. The other part feeds that into a model trained on which content has actually worked on which behavior, and generates new ad content in real time for the screen you are about to pass.

The applicant is Sony Semiconductor Solutions, the division that makes the camera sensors in your phone, which tells you plenty about how the watching would be done.

Nobody browsing a snack aisle has agreed to be a training signal. Retail analytics has crept from counting visitors to profiling paths, and this closes the loop from your footsteps to a message aimed at you before you reach the door.

12. Qualcomm Patents a Way to Give Networks a Peek Inside Encrypted Traffic

Modern internet traffic is a sealed box, and the network that carries it cannot tell a live video call from a background download. Qualcomm has filed for writing a small readable label on the outside of the box.

The filing adds an unencrypted header extension to QUIC, the transport protocol that now carries much of the web. The extension sits outside the encrypted payload and carries lightweight hints about the traffic inside, so routers and base stations can prioritize the call over the download without opening anything. Think of the word fragile written on a sealed package.

For carriers, this is the difference between guessing and knowing, and your video call is the beneficiary.

The tension is that any field readable by every router is readable by everyone else on the path too. Each hint is a little fingerprinting surface, and the fight over which hints are acceptable will happen at the standards bodies that own QUIC, not in Qualcomm’s filing.

13. Samsung Patents Tiny Glass Cuts That Stop Foldable Phone Screens From Cracking

Every time a foldable phone closes, the glass at the hinge bends around a curve that would snap a normal pane, and it does that thousands of times before the warranty runs out. Samsung has filed for a fold zone made of deliberate, engineered cuts.

Thin slits run through the glass parallel to the fold line, separated by solid ribs, filled with a soft polymer, and anchored at each end by small bridges. The slit length is calculated from the glass thickness and the stiffness of the filler, so the zone bends as one tuned system, spreading stress across many small flex points instead of concentrating it along the crease.

It is the same principle as perforated cardboard folding cleanly at the score line, executed in glass at micron scale.

Cuts that save the glass have to stay invisible: a slit pattern is also a pattern of places for optical artifacts, grit, and wear to live. That balance, strength you can feel against texture you cannot see, is the whole game, and it is the standout of Samsung’s 55 filings this week.

14. Google Patent Targets AI That Learns to Fix Account Problems Automatically

You are locked out, you type “I can’t sign in,” and the help page hands you the same FAQ it hands everyone. Google has filed for a support system that reads your account, works out your actual problem, and writes the answer for you specifically.

The system pulls the relevant help document, combines it with your account data, and infers the underlying problem rather than matching keywords. A model then generates a summary tailored to your situation and grades its own output against quality standards before sending it.

The loop is the interesting part: answers that fail the quality gate become training signal, so the system sharpens itself on its own misses without anyone rewriting help articles.

The cost of tailored help is that the helper reads your account history, and a self-retraining support model has to be watched for drift. A generic FAQ wastes your time; a confident wrong answer to a locked-out user does real damage.

15. IBM Patents a System That Catches AI Summaries Making Up Facts

An AI summarizes a long contract, the summary looks clean, and two of its sentences describe terms that appear nowhere in the document. IBM has filed for a checker that flags every summary sentence it cannot trace back to the source.

Both the original document and the summary are converted into embeddings, coordinates in a mathematical space where similar meanings sit close together. The source side covers every combination of sentences, not just single lines. Any summary sentence whose coordinates land far from everything in the source gets flagged as invented, and the mistakes are logged to retrain the model that made them.

It runs in the other direction too, catching what the summary silently dropped. A spell-checker for fabricated facts, wired into an automatic loop.

The soft spot is the threshold: distance in meaning-space blurs exactly where paraphrase shades into fabrication, which in a contract is where the money is. It is IBM’s third entry on this list, and set against the rest of IBM’s 20 filings this week, the theme is unmistakable: machinery for checking the machines.

16. New Google Patent Matches Search Results to Your Intended Format

Type “how to tie a bowline knot” and a wall of articles is the wrong answer; you wanted a video. Google has filed for scoring every query for the format you actually meant.

Each search gets an intent index value per format: video, image, audio, text. The score can come from something as light as a ratio of what similar queries clicked before, or as heavy as a model reading the query’s language. Results are then steered toward the winning format, videos surfacing for the knot, maps and images for “annual rainfall of Brazil.”

It is quiet plumbing, but it is plumbing for the argument between formats that now defines search, where the same answer exists as text, video, and clip.

The honest read is that this is incremental, and that is the story: format steering decides which creators get the click, and a scoring dial this central never stays neutral for long.

17. Google Patents a Search System That Keeps Billion-Item Results Instantly Current

Search for breaking news and the index serving you was built hours ago. Rebuilding a billion-item index every few minutes is unaffordable, so the staleness has been structural. Google has filed for splitting the index into a fast recent layer and a slow bulk layer.

Everything new lands in a small index on fast storage, updated continuously. The bulk lives in a big index on cheap storage, rebuilt on a cycle. Every query checks both and merges the answers, and during each rebuild the newest items exist in both layers at once, so nothing falls through the handoff.

The effect is freshness without the rebuild bill, at the scale where that trade actually hurts.

This is expected infrastructure work, the kind that never makes a keynote, which is why it closes the list rather than opening it. But set against Google’s 21 filings this week, ranking training, format matching, index freshness, you can watch the company re-plumb search one patent at a time.

What ties them together

The interface keeps moving closer to the body. IBM wants your brain states searchable, Meta reads words you only mouth, Sony hides the display inside a sunglass lens, Samsung re-engineers the glass itself, and Apple turns the phone in your pocket into a payment handshake between friends. The machine’s side of the conversation is dissolving into things you wear, and the sensors are arriving faster than the etiquette.

The other column is machines supervising machines. IBM attacks its own safety filters and fact-checks its own summaries, Microsoft fingerprints attack prompts before the model reads them, Nvidia’s models choose their own training diet, and Google’s support system grades and retrains itself. The week’s quiet thread runs between the columns: mouthed words, store footsteps, and hints on encrypted traffic are all signals nobody quite said out loud, and this week Big Tech filed the tools to read every one of them.

Patentlyze reads every new patent application that Apple, Google, Microsoft, Meta, Nvidia, Samsung, OpenAI, Qualcomm, Sony, Intel, AMD, IBM, Adobe, Salesforce, Tesla, Amazon, and Disney file at the USPTO, and writes a plain-English breakdown of the ones that say something about where products are heading. New filings drop every Thursday. The full archive lives in our patent library.