The AI Wire ●

A WIRED investigation published November 2, built on research by the nonprofit AI Forensics and cybersecurity firm Infoblox, found a network of more than 400 connected apps and websites posing as everyday tools — pet-photo editors, language tutors, workout planners — that could secretly load interfaces for making sexually abusive and violent deepfakes from uploaded photos. A server-side switch turned the hidden features on after the apps passed store review, in some cases triggered by arriving through a Meta ad.

The scale ran through both major stores. On Google Play, 27 of the apps were live in September with more than 350,000 combined downloads, including DiffAI at over 100,000 installs; Google had removed most of them before WIRED got in touch. Apple removed 19 apps after WIRED contacted it, including the pet-photo app Viefur, and said the apps complied at review and were changed afterwards — adding it has suspended tens of thousands of apps for similar bait-and-switch tactics in the past year.

Researchers traced the network to Chinese-language codebases and a Hong Kong-registered company, KeLing Technology Limited, registering as many as 30 domain-and-app combinations a week at its peak. Meta said it banned KeLing in 2025 after a cease-and-desist, and reported acting against 11,600 ad accounts, about 4,700 business pages and 209 apps on its advertising platform, while WIRED found more than 84,000 ads tied to the network had run on Meta since January 1.

SemiAnalysis released GDPx on November 2, a benchmark for general agentic capability spanning 11 domains the firm picked for where frontier models will earn their money: chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment and software engineering. Version 1.0 has 206 tasks built with 14 specialist data partners, an eight-hour default time limit, and models running in their native harness — Claude Code for Claude, Codex for OpenAI's models, Antigravity for Gemini. New versions are planned every three to six months, with a public leaderboard at gdpx.semianalysis.com.

The headline number is how low the top score is: Claude Opus 5.5 leads at 12.0%, ahead of Gemini 4 Carbon at 10.9% and GPT-6 Astra at 9.6%. No model leads everywhere — Opus 5.5 wins quant trading, hardware and chip design, Carbon wins biology and edges software engineering, Astra wins recursive self-improvement and materials science — and under binary grading, finance and legal came back at essentially zero for every model. SemiAnalysis also notes Opus 5.5 and Astra are held back in cyber, biology and alignment by their own safety classifiers: strip those three domains out and Astra edges Opus 5.5, while Meta's Muse Wonder Full is rated the best available cyber model. Opus 5.5's lead is not cheap, at $51.41 per task against roughly $17 for Carbon.

Two caveats belong with the headline. GDPx is a paid SemiAnalysis product whose authors want it to become the benchmark of record, and its partner-built tasks are not independently audited. And a 12% top score reflects tasks deliberately chosen to be frontier-hard — it measures how far agents are from full automation of expert work, not whether they are useful today.

OpenAI published LifeVal on November 2, a benchmark that grades AI on everyday responsibilities instead of exams or code: finding childcare, enrolling a child in the right kindergarten, ordering groceries on a budget, managing household finances and getting through government paperwork. The task list was built from the American Time Use Survey, the Multinational Time Use Study and a new survey of 3,995 people on where they most want an assistant's help, across caregiving, health, education, household administration, personal finances, shopping and government services.

The headline score flatters its maker. GPT-6 Astra averaged 80.6% reward, up from 49.3% for GPT-5.2 in December 2025 and 80.0% for GPT-5.6, with example tasks such as organising flood volunteers at 98.0% and diabetes meal prep at 90.4%. OpenAI says it has used LifeVal internally for about a year to track and drive improvement on exactly these tasks.

The limits are structural: OpenAI built the test, chose the tasks, ran the scoring and publishes scores for its own models only, with no independent verification and no opening for third-party models to be graded. An average reward score is also not a share of real-world jobs finished correctly — LifeVal is best read as OpenAI showing which everyday jobs it is optimising ChatGPT for, not as a neutral league table.

OpenAI launched the Grid Protectors Program on November 2, committing the compute campuses it directly develops to reduce electricity use during grid emergencies and extreme weather. The company says it will work with utilities, regulators and policymakers on site-by-site arrangements setting how much demand can be cut, how quickly and for how long, and will publish further commitments as they are finalised.

The pilot is Project Camellia, OpenAI's planned 3,200-megawatt campus in Effingham County, Georgia. Under a 25-year agreement with Georgia Power, up to 1,000 megawatts — about 31% of the campus's peak load, and roughly the average electricity use of 700,000 homes — will be available for curtailment. OpenAI says it will pay the full infrastructure and electric-service costs of serving the site, with contracted power delivered in phases between 2028 and 2032.

The savings figures are the utility's projections, not results: Georgia Power projects about $950 million in annual customer savings from 2029 across its large-load portfolio including this deal, or at least $15 a month for a typical home using 1,000 kilowatt-hours. OpenAI itself concedes the limit of the idea — campus flexibility 'cannot replace investment in new energy supply and transmission' — and the campus, which drew local protests over the summer, is still in development.

OpenAI is expanding the models that power Dots, its always-on agents in ChatGPT. Tibo Sottiaux, who leads Codex, posted late on November 2 that the Dots model family is growing to include GPT-6 Astra and GPT-6.1 Terra, with 'some other models later,' and that Dots will automatically select the model best suited to each task — and can combine multiple models to complete a single request.

Terra is the practical half of the change: Sottiaux positioned it as faster, with a noticeably upgraded personality, and said it 'will help us bring dots to more people.' Dots launched at DevDay on September 29, each with its own cloud computer, and the expansion moves the product from a single-model setup to models split by scenario — which changes both its cost structure and its response speed.

What has not been announced is any change to who gets Dots: OpenAI has published no new plan arrangements alongside the model expansion, and Tibo's post does not mention ChatGPT Plus. Replies to the post ran far ahead of reposts, with many users asking about quotas and availability — questions the announcement itself does not answer.

Users opening the ChatGPT Classic app are now greeted with a retirement date: the app will officially be retired on January 30, 2027, 9to5Mac reported on November 2. The banner tells users to move to the new ChatGPT app and promises their full chat history comes with them. Classic is the old native app, renamed when OpenAI shipped a new Electron-based desktop app that put ChatGPT, Work and Codex under one interface — a change users criticised as confusing, which president Greg Brockman himself called 'kind of a mess.'

Separately, OpenAI is emailing users that it will begin testing a combined Chat and Work experience in the coming weeks, starting with a small group. The merged interface is meant to let users move from exploring an idea to having ChatGPT carry it through in one place — bringing in files and plugins, building a website or presentation, or acting on the web — while Codex remains a separate experience in the desktop app. Brockman had said the company hoped to eliminate the Work tab by the end of the year.

The merge is also where the money changes: as part of the transition, OpenAI plans the shared ChatGPT-and-Codex usage allowance the Wire covered on October 31, replacing separate limits. OpenAI says most users won't need to wait for usage to reset more often than today, that usage will depend on model and task, and that Unlimited Instant chat is included in plans and won't draw from the allowance.

OpenAI announced on November 2 that it is raising its Security Bug Bounty rewards and planning another internal security sprint, this time with outside researchers from Cathedral Systems working alongside its own teams and using OpenAI's models to challenge its defences. Cathedral CEO Gavin Kliger said the collaboration pairs OpenAI's models and security expertise with Cathedral's research to find weaknesses, test fixes and support continuous defence. The sprint is planned, not completed — no start date or targets have been announced.

The money is concrete. A critical vulnerability in a core product such as ChatGPT, Codex or the API Platform now earns up to $30,000; exceptional critical findings remain eligible for up to $100,000 before bonuses, and new multipliers — up to 1.15x for report quality, 1.30x for priority campaigns and a 5–15% loyalty bonus — can lift an exceptional payout to about $172,000. OpenAI says the midpoint of its reward ranges for medium, high and critical findings has more than quadrupled on average.

The new ranges apply generously in transition: reports still awaiting an award will be paid at the new rates, and researchers whose bounties were awarded in the 30 days before the announcement will receive the difference where the new rates pay more. In its previous sprint OpenAI dedicated 25% of its product engineers to finding and fixing weaknesses, using Codex to reproduce findings and draft patches. Jailbreaks and prompt-only safety bypasses remain out of scope for the bounty.

Abu Dhabi signed four agreements with OpenAI on Monday, November 2, after Crown Prince Sheikh Khaled bin Mohamed bin Zayed met Sam Altman and witnessed the signing, the Abu Dhabi Media Office and the Emirates News Agency WAM reported. The deals span government services, energy, investment and education. No financial values or full implementation timetable were disclosed.

The headline initiative is "Machine-Readable Government," led by the Department of Government Enablement: selected government policies and regulations are converted into machine-readable form so AI can help officials with routine compliance checks, including licensing and certification. Its first pilot is already operational at the Abu Dhabi Agriculture and Food Safety Authority, on food imports, running on GPU infrastructure inside the UAE. Alongside it, ADNOC signed with OpenAI and G42's Inception42 — billed as the first energy company to deploy OpenAI's frontier technology — Mubadala signed to use AI for investment analysis and risk assessment, and Inception42 agreed a five-year partnership to spread OpenAI technology across UAE government, large companies and education.

Altman said Abu Dhabi and the UAE had set "an ambitious global benchmark for deploying AI across the public and private sectors." The agreements deepen a relationship that began with G42 in 2023 and produced the Stargate UAE computing cluster announced in May 2025, and they serve Abu Dhabi's stated goal of becoming the world's first fully AI-native government by 2027 — a goal, not a result: food imports are the only live use named so far.

Goldman Sachs has restricted advanced AI coding tools, including Anthropic's Claude Code and Microsoft's GitHub, to its engineers, ending months in which bankers and traders were allowed to experiment with them. Bloomberg reported the change on November 2, citing internal guidance issued in October; traders had used the tools to monitor client trading flows, model market changes and create written and graphic material.

The reversal is notable because Goldman's own leadership promoted exactly that experimentation. In July, Chief Information Officer Marco Argenti praised non-technical staff 'vibe coding' solutions and handing them to one of the firm's 12,000 engineers to refine. Some traders are reported to be frustrated by losing access, while the firm's 'strats' — staff who combine quantitative research, software and trading — keep theirs.

Goldman frames the move as the end of a trial, not a retreat from AI. Spokesperson Nick Carcaterra told Bloomberg the firm makes certain tools available so employees can find the best use cases, then, after assessing productivity and risk management, puts them 'in the hands of those who are best suited to leverage them for the long term.' No incident or audit behind the decision has been made public.

Anthropic filed Amendment No. 1 to its Form S-1 with the SEC on November 2, updating the prospectus it made public on October 26. The amendment adds third-quarter figures: revenue of $17.0 billion in Q3 alone and $33.2 billion for the nine months to September 30, against $2.5 billion a year earlier, with the nine-month operating loss shrinking to $1.3 billion. The share count and price range are still blank.

The filing also fills in a number the first version left empty: about $895.4 billion in purchase and lease commitments, almost all for compute and infrastructure, with $447.5 billion due by the end of 2030 and many contracts on take-or-pay terms. A passage already in the October filing is now being read closely alongside it: the list-price value of Anthropic's internal Claude use grew more than 250 times from January to September, first exceeded monthly personnel expense (excluding stock-based compensation) in April, and Claude wrote over 90% of the company's code contributions in September.

The genuinely new sentence is a risk factor. Anthropic says internal estimates about future models suggest it may imminently meet its Responsible Scaling Policy's definition of dramatically accelerated progress attributable to AI — the point at which AI starts speeding up AI research itself. No current model has been designated past the threshold, the company stresses, and it warns its evidence relies on lagging indicators, so in the past it did not notice a significant acceleration until after it had happened.

Elon Musk has admitted, in public, that xAI's flagship agent does not pay for itself. Asked on X on November 1 whether Grok Bot's weekly usage pool could be reset — after a user reported hitting 100% with four days still to go — Musk replied: "The cost of serving Grok Bot was unsustainably high relative to revenue." He gave no dollar figures, no timeline for relief, and no word on whether prices or limits will change.

The complaints behind the exchange were specific. Paying users reported weekly meters leaping overnight — one widely shared Cursor forum account described usage jumping from about 52% on Saturday night to 100% on Sunday morning, halting paid work mid-update — and neither xAI nor Cursor has said whether the jump was a metering fault, a change in how usage is counted, or both. xAI has never published how large the weekly pool actually is.

Grok Bot, launched on August 11, 2026, is xAI's autonomous agent: it works on its own virtual computer in the background on research, writing, browsing and multi-step jobs. It is not sold separately — it rides along with SuperGrok at $30 a month and up, Cursor plans, and X Premium tiers — and Gwynne Shotwell said in October it had reached roughly 800,000 paid weekly users. Fast growth, Musk's reply suggests, made the compute bill worse, not better.

Sunday, November 1 is Day 28 — the finale of Tibo Sottiaux's 28-day Codex challenge, in which OpenAI promised each day either a clear improvement for Codex and ChatGPT Work users or a full usage reset. Under Tibo's Day 26 post, the final two days were both to be banked resets, and the first is confirmed: Tibo himself posted 'Banked 1/2' at 00:43 UTC on November 1, and reset trackers record a banked credit issued at that time, after monitoring accounts first saw it land on October 31 at 19:59 UTC.

The second banked reset has now landed too. Tibo posted 'Banked 2/2' for Day 28, and TiboReset records the banked grant at 22:32 UTC on November 1, marking the challenge complete at 28 of 28. That closes the run that began on October 5.

A banked reset does not refill usage on its own: it sits on the account as a saved credit until you redeem it under Settings and Usage, and an unredeemed credit can expire. Check your own usage page for the credit rather than relying on tracker counts, which have disagreed on timings before.

OpenAI has added card and device limits to ChatGPT, set out in a new Help Center page. If too many subscriptions are associated with the same payment card or device, ChatGPT may block a new subscription purchase at checkout. Separately, OpenAI now limits how many paid accounts you can use on one device: hit the cap and the account in front of you shows “Device limit reached” and cannot be used there.

The device limit can arrive late — the warning may appear at sign-in, or you may be signed out after a later check, even if the purchase went through and the account worked at first. The limits count ChatGPT Go, Plus and the Pro 100, Pro 200 and Pro 500 plans; Free accounts and Business, Enterprise and Edu workspaces do not count. Signing out of one account does not guarantee another becomes eligible, and buying credits or resetting usage will not clear a card or device restriction. OpenAI frames the limits as privacy and security protections and publishes no numbers for either cap.

The Decoder reads the change as steering users away from stacking several cheap subscriptions on one device toward buying extra credits or usage resets on a single account, which generally cost more — that commercial reading is The Decoder's; OpenAI's stated reason is security. The timing is practical either way: the limits land the same weekend OpenAI announced the shared ChatGPT-and-Codex allowance from December, and days after Codex made account switching easier on mobile.

OpenAI's chief executive spent the weekend in South Korea shoring up the memory supply his data-centre plans depend on. Sam Altman's visit was unannounced: on October 31 he addressed roughly 50 senior Samsung executives at the company's training centre in Yongin, in a speech titled 'The Next 10 Years Changed by AI, Corporate Choices,' and held separate discussions with SK Group chairman Chey Tae-won and executives from SK Hynix and SK Telecom. Samsung confirmed its meeting; SK Hynix told CNBC it could not confirm the lunch.

No new supply agreement came out of the trip. The meetings build on the October 2025 Stargate letters of intent under which Samsung and SK Hynix agreed to supply memory for OpenAI's infrastructure, including a production target equivalent to as many as 900,000 DRAM wafer starts per month — a target, not a binding order. What has changed since is bargaining power: Samsung and SK Hynix together control most of the world's high-bandwidth memory, Samsung has said its 2027 HBM supply is sold out, and SK Hynix has said production through 2029 is effectively accounted for.

The pressure behind the trip is cost as much as volume. Research firm TrendForce forecasts HBM's blended average price could rise 121% in 2027, and the visit landed in the same news cycle as reports that OpenAI has cut its current fundraising target from $30 billion to about $20 billion — a figure OpenAI has not confirmed, and which the Wire tracks separately as an unresolved rumor. Neither Samsung nor SK disclosed any new investment commitment or pricing terms from the weekend.

The National AI Safety & Privacy Standard that Microsoft developed with the American Federation of Teachers and the United Federation of Teachers took effect on November 1. Announced on September 9, the 31-page standard is not a law: any US school district can add it to a new or existing Microsoft agreement without renegotiating the whole contract, and once added its terms are binding contract language rather than a vendor promise on a webpage.

The substance is specific. Student and educator data may not be used to train AI models, sold, or used for advertising; AI companion features designed to foster emotional attachment or keep students engaged past the learning task are banned; consequential decisions require human oversight; providers must explain their products in plain language, report breaches within 72 hours, and submit to third-party audits. Districts that adopt the standard gain contractual remedies, including termination and damages.

So far it binds one company. The agreement is structured as a memorandum between Microsoft and the AFT-affiliated National Academy for AI Instruction, with signature slots left open for other providers, and the teachers' unions say OpenAI and Anthropic have expressed willingness to sign. It says nothing about whether classrooms should use AI at all — the UFT's president has said AI use in grades K-8 is not appropriate right now — only the terms for districts that do.