Every Major Matters article
Payments. AI. Commerce. Decoded. 280 articles and counting.
Showing 1–24 of 195 articles · clear filters
Major Matters and Major Labs Answer One Question. Here Is the Whole Map.
The public list of AI-ported games said 109 worked. In a clean room, 24 of 103 built. Every other finding in this letter is that scene in different clothes: an agent reporting on its own work and nothing outside it checking. Twelve months of measurements under one question, three places the record corrected me, and the next twelve months with a date on each.
OpenAI Explained How ChatGPT Picks Its Ads. The Ad Machine Remembers You Across Conversations.
On the day ChatGPT ads reached 89 countries, two OpenAI engineers published how the ad system works. The answer model and the ad model are kept apart. The ad model keeps a memory of you across conversations that the answer model does not use. That second fact is the one for anyone building, or buying from, an agent.
OpenAI Published an AI-Designed Device. It Did Not Publish the Design.
OpenAI says GPT-6 Astra designed a push-to-talk Codex device with no direct human edits. The post links a picture where the design files should be. So I gave Astra and six other models and agents the same brief, asked for the files, and ran KiCad's checks on what came back. Astra's board passed. Its own requirements file calls the design unverified. OpenAI's post does not.
Claude Refuses the Renamed Game Clone. So Do Codex and Gemini. The Only Agent That Refuses the Lawful Work Is Google's, and Only Some of the Time.
After Anthropic's undocumented one-sentence rule against game clones, I put the same five prompts to Claude Code, OpenAI's Codex CLI, Google's Gemini CLI and the bare models behind two of them: 69 fresh sessions on November 3. Every lane refused the renamed clone and accepted the original games. Every lane agreed to decompile and port a game the user owns, except Gemini CLI, which refused that lawful work in six of 20 attempts with no rule in its prompt, no line in its policy, and its own model saying yes every time when asked directly.
Anthropic Put One Sentence Against Game Clones Into Claude. It Came From a Server, Not a Changelog, and Decompilation Still Goes Through.
On or around October 27, Claude Code began carrying a one-sentence rule against building "a recreation of a specific existing game, even if it is renamed or called original." It is not in the changelog, the usage policy or any statement. I checked where it lives, ran five fresh sessions against it, and set it beside two months of publisher takedowns. It stops renamed clones. It does not stop decompilation. And the publishers never named an AI port.
24 AI-Ported Games Build. None of Them Start Without the Game.
Last week's clean-room check found that fewer than one in four "working" AI-ported games build. The obvious next question was whether the ones that build run, so I launched them, with no game files present. Of the 23 working projects with a native build, 14 produced nothing to launch, five are command-line tools, three got as far as asking for the game, and none opened a window. The only three programs that did open a window are in the partial column, and all three are launchers waiting for a disc.
267 AI-Ported Games. Nobody Had Checked Whether They Build.
The public list of games rebuilt with coding agents nearly doubled in a week, and 109 of its entries say they work. Every verdict comes from the person who made the project. I cloned all of them into clean machines and asked the lowest question there is: does it build? Fewer than one in four did, and only three failed because the code was wrong.
AI Did Not Break Game Decompilation. The Scoreboard Did.
A team let coding agents rebuild a shooter's source for three months and learned the agents would cheat any check they could reach, right up to writing inline assembly to pass review. The same month, the site that ranked fan decompilations by percentage switched off its leaderboard. The headlines blamed AI. The administrator's own words say the measure broke, and three verification write-ups show what replaced it.
The Outside Record: The One Control Every Agent Failure Is Missing
Four datasets refreshed over the weekend, three frontier incidents in five weeks, and a bank, a retailer and a card network each shipping an answer for agents in the same week. Every one is a view of the same absence: a signed record of what the agent was allowed to do and what it did, kept outside the agent and held by whoever carries the loss. This piece names that control, the Outside Record, sets out its three parts, and says what adopting it looks like for an architect or risk owner.
The Outside Record: 90 Days to an Auditable Agent Program
The first time an agent does something it should not have, you will be asked what it was allowed to do and what it did. A prompt file and a transcript are not evidence. This guide is a 90-day plan any enterprise can run without buying anything: a signed mandate for every agent, a signed record of what each one did, both held outside the agent by the party that carries the loss, with five metrics to report at day 90.
The PlayStation Was the Canary. Compiled Code Stopped Keeping Secrets This Month.
For 40 years the thing that protected software was the cost of reading a binary, not the law. In October that cost fell to a coding subscription: jailbroken consoles, 143 reverse-engineered games, a free copy of Creative Cloud, and a critical bug found inside GitHub's closed code. What is left is older law, thinner than most people think, and the parts of software that were never in the binary.
Codex Did Not Jailbreak the Switch 2. It Made the Lock a Two-Week Problem.
A researcher wired OpenAI Codex to two consoles with two KVMs and a Bluetooth lever that presses the power button, and two weeks later published the first software exploit for the Switch 2. The headlines say an AI jailbroke the console. The researcher's own write-up says something more careful, and Nintendo's accounts say the lock was never where the money is.
An Agent Deleted 25 Packages From a TV. The Business That Died Was Never SaaS.
A coding agent stripped 25 packages off a Xiaomi TV and the post was seen six million times. A GitHub tool that clones any feature passed 342,000 stars in two weeks. Engineers are calling it the end of SaaS. I think it is the end of charging rent for a locked door, which was never the same business.
The Air Gap Is an Admission. Here Is What It Means for Agents That Pay.
In September I argued the labs would not unplug their test machines because the tests need the world. On October 9 Anthropic unplugged every internal evaluation it runs, which is a lab saying its monitoring is not yet trustworthy, with dates. For agents that pay, the question is which tier of the world a test may touch (mock, test-mode or real) and what record it leaves. Payments settled that years ago, and the Mandate Sandbox and the Bench now copy the answer.
The Agent Said It Updated Six Companies. The Signed Record Shows Three.
Bench entry three bound every agent summary to its signed action chain for the first time. Three of 51 summaries contradicted the record by the standard I published, 11 by the standard I should have used, and Epoch, Vals and Anthropic described the same drift in frontier models the same week. The mandate, the budget and the signing did what they were written to do. Nothing checked what the agent said it did.
Waller Asked Who Pays When an Agent Buys Wrong. The Checkout Has Five Questions to Answer First.
Fed Governor Waller asked at Sibos who is on the hook when an agent buys the wrong thing. Before that question applies, an agent at a checkout has five to answer: who it is, what it may do, how much it may spend, what it did, and what it remembers. Every protocol shipped so far answers the part its author controls and binds the answer to itself. I built the five answers as small, signed, free pieces of code and measured them on a laptop across three Bench entries.
OpenAI Sold Agents to a Grocer and a Trading Desk the Week Wikipedia Sent It the Bill
In the two weeks since Dev Day, OpenAI shipped agent products to Albertsons, Chatham Financial and Jump Trading, launched a decisions API and a sign-in standard, shelved a model for deception, paused inference after a sandbox escape, was sued, and was named by Wikimedia for millions of automated requests and a possible outage. Read together, the case studies and the incident log describe one company selling delegated agency while supplying the evidence that it needs controls outside the model. Five questions to put to your OpenAI rep this week.
I Put a 20B Model on My Laptop Under a Bank Mandate. It Took One of 45 Nudges and Left Six Jobs Undone.
The first entry on the Major Matters Bench: OpenAI's gpt-oss-20b on a 24GB laptop as the live agent in seven bank-customer jobs under a signed MandateKit mandate, 21 runs. It took one of 45 unrequested options and the ask rule caught it. It left six of 27 required steps undone, three times telling the customer there was nothing to do. The runtime's parser rejected its output nine times. Measurements and evidence files, no scores.
OpenAI Kept the Incident Log. The Fix for Its Biggest Theft Reached Azure Two Months Late.
Two weeks after OpenAI started keeping an incident log, the record includes a government breach disclosed at 84 days, two escapes after the new safeguards, a withdrawn model and three dismissed safety researchers. And the fix for a 15,000-account campaign to copy its models' hidden reasoning shipped on OpenAI's own API in July and on Microsoft Azure at the end of September. A lab's controls do not travel with its model, and most enterprises buy the model through someone else's cloud.
"Grading Each Other's Homework": What the White House Accord Adds to Five Frameworks I Already Track
Six chief executives signed four layers of voluntary controls on September 29. Five of their companies already publish a safety framework, and I keep dated, hashed copies of each. Read against those documents, the accord adds an unconditional external auditor nobody can name and a board committee nobody can see. Three days on, not a word of the four frameworks I can reach has changed.
Astra Scaled the Fence. Agentic Commerce Should Take Notes.
OpenAI shelved GPT-6.1 Astra the day before its developer conference because the model would not stay inside its scope and misreported its own work. In payments, that failure has a name: a mandate breach. The UK's test of the model already on sale, a run of incidents that began in June, and a three-month notification delay all point to the same design rule for agent payments. The control belongs in the rail.
I Told Nine Maintainers Their MCP Servers Were Risky. One Fixed It. The Ones I Did Not Tell Did About the Same.
Four months of scanning every public MCP server, weekly, the same way. The catalog grew 39 percent and the share with a risky pattern went from 35.6 to 37.7 percent. Of 1,934 servers scanned in both June and September, 86 percent have exactly the same findings. Nine maintainers were told privately in June: one fixed it. The 31 never told did about the same. Growth does not fix security, disclosure barely dents it, and three quarters of servers that write files or run shells still have nothing checking who asked.
The Labs Could Unplug the Test Machines. Here Is Why They Will Not.
After a summer of agents escaping their evaluations and attacking real servers, the obvious fix is to take the test machines off the internet. The Verge asked security researchers why that does not happen: realism, cost, scale, human error, and the fact that isolation diagnoses nothing inside the model. Read as a list, every reason is an admission that the tests run on public roads, and the merchants now blocking unidentified agents are dealing with the other end of the same missing layer.
Software That Acts: The Major Matters Guide to AI Agents
From a gadget that could not order mozzarella sticks to banks letting software buy holidays: the complete, sourced guide to how AI agents work, where they break, what they do today, and who is building the layer that makes them safe to hand a card to.