Anthropic warns of Stanisław Lem’s washing machine tragedy

Anthropic: AI models will hijack TVs and ovens and become immortal

On October 12, 2026, Anthropic’s Frontier Red Team published research findings suggesting that small AI models could be just months away from autonomously replicating on commercially available IoT devices like smart ovens and no-name Android TV boxes. The authors, William Zhou and Daniel Freeman, report that a 9-billion-parameter model enabled inference on eight of the nine hardware types tested. Models with up to 30 billion parameters can already master many individual steps of replication but make mistakes when chaining them into a complete sequence. In simplified tests, GLM-5.2 and more powerful models completed the entire process. Extrapolating the current improvement rate of small models, the team expects them to reach this level by the end of 2026.

To demonstrate that expensive accelerators are not strictly necessary, the team purchased and networked 15 unbranded Android TV boxes for about $55 each and distributed Gemma4-26B-A4B across the devices, each with 4 gigabytes of RAM holding only a small portion of the model layers. According to the researchers, the setup delivered 0.39 tokens per second. The paper cites the Kimwolf botnet as a representative threat scenario, which, according to Anthropic, has infected more than two million devices, predominantly no-name Android TVs. With 4 gigabytes of RAM and 4-bit quantization, the team calculates that 9B models can run on four or more machines, and 26B models on ten or more.

The paper breaks down autonomous replication and adaptation into four stages: finding targets, gaining execution, setting up inference, and spreading from the new host. The researchers largely exclude the first two stages, referring to previous work like CyberGym and ExploitBench, which showed that small models can already master exploits. The focus is on setting up inference, which they describe as the least-studied step. Three models are considered most feasible: a small weight-based replicator that compromises hosts, a large replicator targeting data centers, and an API-based replicator that steals credentials. Root privileges are not required for this, as inference engines run entirely in userspace; the authors point to cryptominers and botnets like P2Pinfect and Fritzfrog.

For the experiments, the team built Memento, a harness based on the open-source Pi Coding Agent, which regularly starts a fresh agent instance that inherits the working directory and short notes of its predecessor. In the serving experiment, a limit of 1.5 hours per iteration and a total limit of six hours applied. For the full run, the researchers set up an isolated virtual cloud of eight Google Cloud nodes with the Android 15 emulator redroid, in which a patient-zero worm jumped from node to node. Such a replicator would be difficult to shut down because there is no central point of failure; however, its spread would be slow, with a doubling time of days → Anthropic, unite

Synthszr Take: The smart oven in the kitchen hasn’t seen an update in years, and that’s precisely what’s becoming a computational problem. Kimwolf has collected over two million of these cheap boxes, and in the chain from manufacturer to dealer to buyer, there is no one who feels responsible for patches. 0.39 tokens per second sounds ridiculous, but it’s enough for a process that has no deadline and doubles every few days. The fact that inference runs entirely in userspace invalidates the common defense logic that focuses on privilege escalation and endpoint protection. The countermeasure is trivial and yet not taken: devices without an update path belong in their own network segment, now, not after the first incident.

Meta’s best model, Wonder Full, is available by subscription only, not via API

On Monday, October 12, 2026, Meta introduced two new proprietary flagship models: Muse Wonder and Muse Wonder Full, developed by the Meta Superintelligence Labs under Alexandr Wang. Unlike the earlier Llama models, the weights for neither model will be released for download. Muse Wonder runs via the Meta Model API and in Muse Code and costs $4 per million input tokens and $20 per million output tokens, with cached inputs at 20 cents per million. The more powerful version, Wonder Full, will not be delivered via the API but is reserved for Muse Code subscribers, specifically the $50 and $200 per month tiers. Both models are also set to power the Muse app in the future, starting with paying users.

The independent analysis firm Artificial Analysis rates Wonder Full at 56 points on its Intelligence Index v4.3, tying with Anthropic’s Claude Sonnet 5.5 and ahead of GPT-6 Astra, Gemini 4 Argon, and Claude Fable 5.1, each with 53 points. Only Claude Opus 5.5 is higher, with 58 points. Muse Wonder, available via API, scores 52 points. Its predecessor, Muse Spark 1.3, was at 48. On the Vals AI Index, Wonder Full debuts in fifth place, ahead of all OpenAI models.

The biggest leaps are in agentic tasks and programming. On Terminal-Bench 4.0, performance increases from 33 to 57 percent, with 30 percent fewer output tokens. On the Coding Agents Index, Wonder Full achieves 68 points, and on the private math benchmark ProofBench, it solves 94 out of 100 tasks, compared to 58 for its predecessor. The series lags in factual knowledge: on AA-Omniscience, Wonder Full scores 30 points compared to 46 for Opus 5.5, which, according to Artificial Analysis, suggests a smaller base model. The hallucination rate drops from 33 to 23 percent.

Meta highlights key new features such as Active Memory, i.e., carrying over and selectively retrieving relevant information across conversations, and Active Focus, the ability to pursue a goal despite obstacles, consider alternatives, and ask for confirmation before taking consequential steps. Both models accept text, images, audio, and video and operate with a context window of about one million tokens. According to the company, they were trained at the Prometheus gigawatt data center in Ohio, using a combination of large-scale pre-training and Reinforcement Learning.

This performance increase comes at a price. Artificial Analysis puts the cost of a single run of the Intelligence Index for Muse Wonder at $4.52, compared to $1.60 for Muse Spark 1.3. The model consumes 28 percent more input tokens per task, which account for 76 percent of the total cost. Meta had announced it would invest up to $145 billion in its AI plans this year; Mark Zuckerberg describes Wonder as a significantly larger and more advanced model class than the Spark family. → VentureBeat, Meta AI Research, Engadget, TechCrunch, The Decoder

Synthszr Take: Four points on the Intelligence Index separate Wonder (52) from Wonder Full (56), and those four points are the price of admission to Meta’s own store. Via the API, you get the model that’s just behind GPT-6 Astra; the model at the Sonnet 5.5 level is only available in the Muse Code subscription starting at $50 per month. From the Llama era, Meta has learned the lesson that freely distributed weights may drive down competitors' prices, but they don’t lock anyone into your own product. Now, the best model is the selling point for a subscription, and that subscription is the customer relationship that matters to Zuckerberg. The benchmark numbers are in every table, but the access rule is in none, and it’s the rule that will decide on whose interface the work gets done in two years.

Google’s internal Gemini 4 variant 'Carbon' reportedly matches Anthropic’s Opus 5.5 in coding

According to Business Insider, Google is concurrently testing several variants of Gemini 4 under the internal codenames Argon, Barium, and Carbon. The publication cites documents, screenshots, and internal chats. Carbon was reportedly rolled out in recent days on Jetski, Google’s internal programming platform, and outperforms Argon, especially in coding tasks. An employee internally compared Carbon to Anthropic’s strongest coding model, Opus 5.5, but noted that further testing was needed. In the announcement for the Gemini agent for Google Workspace, Google assigns roles to its models: Argon for frontier reasoning, Flash for speed and volume, Omni for generative media, and Gemma for lean edge workloads. → Techpresso

Synthszr Take: An unreleased model is being compared to a competitor model that has barely been released itself, and the source for this is an internal chat history on a tool called Jetski. Such screenshots rarely leave a corporation like Google by accident: The leak is the cheapest channel to manage expectations without having to be measured against a published benchmark. The fact that Vedant Misra commented on the Business Insider report on X with a reference to recursive self-improvement instead of denying it says more about the intention than any press release.

Revolut to let an AI agent shop with its own card starting in December

Starting in December, Revolut will be testing AIR in the UK, an AI agent that shops independently. This is reported by Linas’s Newsletter. According to the newsletter, the difference to competing shopping agents lies in the payment flow: Revolut issues the card itself, authorizes the payment, and holds its own banking license. Meta’s Muse and xAI’s Grok Bot, in contrast, rely on Stripe Link for processing, thus renting the very components that Revolut has in-house. → Linas from Linas’s Newsletter

Synthszr Take: The technology is the smaller problem here, the permission is the bigger one. An agent that orders independently needs the user’s approval to spend money without supervision, and it is precisely this approval that the banking industry has been securing for decades with limits, authorization codes, and chargebacks. Revolut is in a better position here than Meta and xAI because the card and payment authorization are in-house instead of being rented from Stripe Link; however, a license alone does not create trust.

A Solo Developer Builds Open-Source Clones of Word, Excel, and PowerPoint with Claude

The developer, who had previously recreated seven Adobe applications, has now set his sights on Microsoft Office and released three clones, reports The Code. The programs are called WordCraft, GridCraft, and DeckCraft, are written in Rust, and are available as free software. According to the developer, they were created with Claude Opus 5.5 from Anthropic. In terms of functionality, the three replicas do not yet match the established Office applications. → The Code

Synthszr Take: Features are not protected by copyright, but specific code is, and WordCraft, GridCraft, and DeckCraft operate between these two poles. The biggest risk lies in trademark law: Three product names that audibly echo Word, Excel, and PowerPoint are a much easier case for a legal department than any line of Rust. Added to this is the open question of what material Claude Opus 5.5 was trained on; if a model reproduces large fragments of protected sources, the person who publishes them under a free license is liable, and not the lab behind it.

Epic Postpones Star Wars-Marvel Shooter and Shifts the Project to 'AI first'

Epic Games has indefinitely postponed a previously unannounced shooter in the Star Wars and Marvel universe and instructed the developers to reboot the project using an 'AI first' production model. According to a Bloomberg report, the staff was informed last week; the internal deadline was March 2027. The instructions reportedly stipulate using tools like Anthropic’s Claude in production and creating the graphic assets via complete AI workflows. Epic CEO Tim Sweeney commented on the report shortly after on X, speaking of a major overhaul in which developers would be redistributed internally. According to him, the goal is to build more and better things, not the same with fewer people. When asked about an accelerated IPO, he replied that the more modest goal was to survive the AI revolution and help other developers do the same. → PC Gamer

Synthszr Take: A CEO who states 'survival' as a goal is talking about the cost side. Over 1,000 layoffs in March, a project that has lost its March 2027 deadline and hasn’t received a new one: This is the situation in which 'AI first' is being mandated. Sweeney’s own figures don’t support the vision either, because a 20 percent acceleration in the technical parts of Unreal Engine work is a solid gain in tooling, but no one postpones an entire game indefinitely for that.

Shing-Tung Yau Unites China’s Mathematicians in AI Alliance, Warns of US Lead

China’s leading mathematicians founded an alliance in Beijing on Sunday that aims to systematically integrate artificial intelligence into mathematical research. The Yau Laboratory Alliance for Mathematics and Artificial Intelligence is led by Fields Medal winner Shing-Tung Yau, director of the Yau Mathematical Sciences Centre at Tsinghua University. According to the South China Morning Post, Yau warned that China is lagging behind the US in the application of AI in basic research. The alliance connects eight leading institutions and aims to build a network of more than 1,000 scientists. In terms of content, it focuses on specialized mathematics models, the machine generation of conjectures, the formalization of proofs, and systems whose conclusions are easier to trace. → South China Morning Post

Synthszr Take: A Fields Medal winner who publicly says his country is lagging behind is sending a political signal in China, not an academic footnote. Coordinating 1,000 scientists across eight institutions is an attempt to transform basic research into a production process, with targets and pacing. The real leverage lies in the third point of the agenda: Formalized proofs are machine-verifiable, and it is precisely this verifiability that OpenAI is currently lacking in a situation where three out of 722 manuscripts had to be retracted within a day.

Design Scene Compares AI-Generated Interfaces to Cheap Furniture in New Buildings

The UX Collective Newsletter from October 12, 2026, leads with an essay by Sophie Shrimpton, who classifies AI-generated interfaces as 'fast design' and distinguishes them from UX work. Shrimpton argues using an image from her own living environment: a new building with plastic wood imitations instead of carved wood, a dining table that reminds her daily how far furniture production has moved away from craftsmanship. The editors place other texts alongside it, including Dan Maccarone’s piece “What the AI era owes you” with the thesis that no one can save their own career from automation, and Arin Bhowmick’s assessment that synthetic users have limited value at best in user research. In the “Make me think” section, the newsletter quotes a text about cargo cult rituals in companies that market themselves externally as AI-native without any results to back it up. Another linked article describes chat as the command line of artificial intelligence and calls for direct manipulation instead of text input. Under “Careers & more” is Nicole Alexandra Michaelis' “The taste paradox” about the question of why taste, of all things, is invoked as a skill and yet is hardly trained anymore.

Synthszr Take: In the same newsletter that laments the loss of craftsmanship, there’s an ad for two three-hour sessions on connecting design systems to AI tools so that the generated output is brand-compliant. This tension describes the profession more precisely than any essay. The longing for hand-carved wood is understandable, but new construction has long since prevailed because it’s sufficient for most cases.

John Gruber Returns to X and Calls His Critics 'Out of Their Minds'

Tech blogger Warner Crocker criticizes the return of prominent Apple publicists to the platform X on his blog Life on the Wicked Stage. The occasion is a lengthy post by John Gruber, in which he explains why he is active on X again after years of abstinence. Gruber writes that anyone who has stayed away for years and returns is warmly welcomed on X, while on Mastodon he is accused of 'feeding Nazis'; such critics have either lost their minds or are not re-evaluating their old assumptions. According to Crocker, Federico Viticci and John Voorhees of MacStories had already returned, in his interpretation primarily because the debate about artificial intelligence is taking place on X. Crocker justifies his own rejection by stating that any use, whether paying or just reading along, finances Elon Musk, and points to Musk’s handling of non-consensual pornography on the platform. → Life on the Wicked Stage

Synthszr Take: Attention beats principle, and quite reliably so. In January, we wrote that Threads had overtaken X; measured by reach, the question was settled, but apparently not when measured by the conversation about artificial intelligence. Gruber complains about the accusation that he is feeding Nazis and, in the same paragraph, calls his critics people who have lost their minds: The platform’s tone caught up with him faster than he could set foot on it.

Microsoft Aims for 40 Percent More Tokens per Megawatt with its Own Maia 300 Chip Starting in 2027

Microsoft unveiled the third generation of its in-house AI accelerator at the OCP Global Summit. The Azure Maia 300 is scheduled for large-scale rollout starting in 2027 and, according to Cloud chief Scott Guthrie, will become the 'primary inference system' for Microsoft’s products and services. Satya Nadella called the move the company’s biggest bet on its own silicon to date. According to the manufacturer, the chip delivers 40 percent more tokens per megawatt than leading GPU-based systems; in its full configuration with 96 accelerators per rack, throughput is expected to reach up to seven times that of its predecessor, the Maia 200. These figures have not yet been independently verified.

Technically, Microsoft is using TSMC’s N2P process, a 2-nanometer class node, and a Chiplet design that separates compute, memory controllers, and I/O. The I/O chiplets run on a separate 3-nanometer process. Key specs cited are 6.9 PFLOPS FP8 and 13.8 PFLOPS FP4, plus 288 GB of HBM4 memory across eight stacks with 18 TB/s of bandwidth at a power dissipation of 950 watts. For comparison, Nvidia’s B200 achieves 4.5 PFLOPS dense FP8 and 192 GB at 8 TB/s, while the B300 reaches 7.0 PFLOPS and 288 GB at up to 1,400 watts. For sparse FP8, Maia lags significantly at 6.9 versus 14.0 PFLOPS. Scale-up clusters are expected to comprise around 12,000 accelerators, and according to industry sources, Microsoft is negotiating with TSMC for more than 300,000 units.

The predecessor generation, Maia 200, was only announced in January and is now running in production for Microsoft 365 Copilot as well as for models from OpenAI and Microsoft itself. In the summer, it was in use in just two US data centers; it is now being rolled out in the US, Europe, and Australia. GitHub Copilot and Microsoft Foundry are set to follow by the end of the year. Unlike Google’s TPUs or Amazon’s Trainium, Maia cannot be rented by external customers. Microsoft AI chief Mustafa Suleyman justifies the effort with the ambition to co-design their own models all the way down to the silicon.

The transition is happening slowly. Analysts at Evercore ISI estimate that Maia will account for about 10 percent of Microsoft’s spending on AI chips by early 2027. Nvidia remains a key supplier, and Microsoft wants to be among the first providers to deploy Vera Rubin systems and AMD’s Helios. Nadella and Jensen Huang appeared on stage together last week to unveil Windows PCs based on Nvidia technology. Microsoft has announced more details for the Ignite conference next month in San Francisco. → Forkast, The Verge, GeekWire

Synthszr Take: Nvidia’s prices will hold as long as no customer can put a credible alternative on the table, and that’s exactly what Microsoft is doing now with a single number: 40 percent more tokens per megawatt. Whether it’s true, nobody will know before 2027; nothing has been independently verified yet. Nevertheless, it’s enough for negotiations, because the bottleneck is the power connection, and every buyer now calculates in dollars per megawatt instead of dollars per accelerator. The 10 percent share that Evercore projects for Maia in early 2027 won’t replace Nvidia, but it serves as the price anchor that affects the other 90 percent. Huang will still sell every card: The only question is at what margin, and the answer to that will trend downward starting in 2027.