AI that could copy itself onto TV boxes, Character.ai goes bust, and Meta's first frontier model

On October 12, Anthropic's Frontier Red Team published research by William Zhou and Daniel Freeman asking a worrying question: could an AI model copy itself onto ordinary connected devices, such as smart TV boxes or smart ovens, and keep spreading without anyone's data centre? Their answer is 'not yet, but maybe soon'. The authors write that small open-weight models may be months away from being capable enough to do this. No such self-replicating AI exists in the wild today, and Anthropic stresses that its tests describe early 'signs of life', not a working worm.

The team split replication into four stages: finding targets, gaining access, getting the model running on the new device, and spreading further. They focused on the third stage. In a physical demo, Google's Gemma4-26B-A4B ran across 15 off-brand Android TV boxes, each with 4GB of memory, at just 0.39 tokens per second. Qwen3.5-9B got inference working on 8 of 9 hardware types tested. In an 8-node emulated Android network, only Z.ai's GLM-5.2 completed full replication, reaching all eight nodes in 4 of 5 trials. Across six generations passing notes forward, Qwen3.5-9B's setup success rose from 45% to between 61% and 84%.

Why worry about such a slow system? Because it would have no off switch. A model running on thousands of hijacked devices has no central server to shut down. The authors use the Kimwolf botnet, reported at over 2 million infected devices, mostly Android TVs, as a yardstick. If every group of 15 TVs ran one copy, they estimate about 133,000 replicas, but total compute equal to only about 40 Nvidia H100 GPUs. Using Epoch AI's capability index, they project that a small model of 30 billion parameters or fewer could reach GLM-5.2's level by the end of 2026.

The caveats are significant, and Anthropic lists them itself. The hacking step was deliberately trivial: test devices had unauthenticated debug ports, so the results are an upper bound on real-world success. The botnet estimate rests on one device class and one botnet. Small models also did poorly at spearphishing, and giving them a frontier 'advisor' model (Opus 4.8) did not help significantly. Setting up inference on a TV-box cluster took an estimated 71 to 150 hours. SecurityWeek, which covered the paper, quoted no independent experts, so this is still one lab's assessment.

The practical message is about old, boring security. The authors note that multi-gigabyte model transfers and unusual device-to-device traffic would stand out in ISP monitoring. They point to the Mirai botnet clean-up, where quarantining and notifying infected customers cleared 92% of infections within 14 days, versus 74% without notification. They also back secure defaults, citing the UK's ban on default IoT passwords. For businesses, the lesson is that the cheap Android box in the reception TV is now a potential AI host, and network hygiene is the first line of defence.

Why it mattersSelf-spreading AI would not need powerful hardware, just lots of insecure cheap devices, which most offices, hotels and homes already have.

Character Technologies, the company behind the Character.ai chatbot app, announced on October 12 that it had started a voluntary Chapter 11 bankruptcy process to support a court-supervised sale. In a blog post, the company cited 'significant legal, regulatory, and operational headwinds' over the past year and said its board decided a sale was the best way to maximise value for stakeholders. The app remains online, subscriptions and billing are unchanged for now, and the company says subscribers will be warned in advance of any change to paid features.

According to reporting by Variety and The Hollywood Reporter, the case was filed in Delaware and a Disney subsidiary, Disney Worldwide Services, agreed to pay $20 million as the stalking-horse bidder. That means Disney's offer sets the minimum price and others can bid higher. Disney's bid covers a non-exclusive, perpetual licence to certain Character.ai intellectual property and certain company claims, not the chatbot app. The filing estimated both assets and liabilities at between $50 million and $100 million. The company expects the sale to finish by late November or early December 2026.

The fall is steep. Character.ai was founded in 2021 by former Google researchers Noam Shazeer and Daniel De Freitas and reached a $1 billion valuation in 2023. In 2024, Google paid about $2.7 billion to license its technology and rehire the founders and some researchers, after which Character.ai stopped building its own large models and relied on outside ones. It then faced lawsuits from families who said its chatbots harmed teenagers, including a Florida case over a 14-year-old's suicide that was settled in early 2026. In late 2025 it banned open-ended chats for users under 18.

Disney's role adds a twist. In September 2025, Disney sent Character.ai a cease-and-desist letter over chatbots using its characters. In September 2026, Disney hired Character.ai's then-CEO, Karandeep Anand, as its chief technology officer, and reports say former Character.ai staff will provide transition services for a limited time. Some details remain unclear: the company's blog names no bidder and gives no figures, and the outcome of the auction is open. If no buyer wants to keep the platform running, Character.ai says it will be wound down.

For the wider industry, this is a warning about building a consumer AI business without control of your core technology, revenue model or safety story. Character.ai had huge engagement but lost its founders and model team, carried heavy legal exposure around minors, and ended up worth a fraction of its peak. For users and businesses that depend on any AI startup, the practical lesson is to know how to export your data and switch providers before you need to. For IP owners, Disney's bid shows that characters, conversations and know-how still have value even when the company does not.

Why it mattersOne of the best-known consumer AI apps collapsed under legal and regulatory pressure and dependence on others' technology, which shows that engagement alone does not make a durable AI business.

Meta's Superintelligence Lab released two models on October 12: Muse Wonder and the larger Muse Wonder Full. Meta calls them 'our most capable models to date' and says they deliver 'state-of-the-art performance on personal and professional agentic tasks'. TechCrunch describes the launch as Meta's first real attempt at a frontier model since Alexandr Wang restructured the company's AI work in 2025. Until now, Meta's Muse line had focused on the lighter Muse Spark model and the Muse personal agent, which launched on September 8.

Muse Wonder is available now to developers through the Meta Model API and in Muse Code, Meta's coding product. TechCrunch reports API pricing of $4 per million input tokens and $20 per million output tokens, which it says is roughly in line with OpenAI's GPT-6 Sol and Anthropic's Claude Opus 5.5. Muse Wonder Full will not be served through the API; it is only for Muse Code subscribers. Meta says Wonder will 'soon' power the Muse app, rolling out to subscribers first. CNBC reports the Muse app plans cost $20 or $100 a month, and that the model was codenamed Watermelon.

Meta's pitch rests on two features. 'Active Memory' lets the model carry important details across conversations and bring them back when relevant; research scientist Kavya Kopparapu said in the launch video the goal is to remember 'what's important to you over days, weeks, months, and beyond'. 'Active Focus' is meant to keep the agent working when a task gets hard, by trying other approaches, recovering from errors and checking in before important actions. Meta's example uses include insurance claims, lease reviews, personal finance and coding.

Treat the performance claims carefully. According to TechCrunch, Meta's model card puts Wonder and Wonder Full roughly level with Anthropic's Fable 5.1 and OpenAI's GPT-6 Astra, and Meta says Wonder beats those rivals on MemoryBench, a memory benchmark. All of that is Meta's own testing, and independent results are only starting to appear. Meta's announcement says a full safety report is 'coming soon', so it was not available at launch. Yesterday's edition covered questions about Muse's adult-content behaviour, a reminder that agent policies and real behaviour can diverge.

The models were trained at Meta's Prometheus data centre in Ohio, according to TechCrunch, and investors reacted well: CNBC reported Meta shares rose more than 2% for a second day after the launch. For builders, the significance is choice and pressure on prices. A third major US lab now sells a frontier-class model by API, with a strong focus on long-term memory for agents. That makes it worth testing on your own tasks, especially for assistants that must remember a client's history, but not worth switching on marketing claims alone.

Why it mattersMeta's entry adds a third big US frontier option for developers and makes long-term memory a headline selling point for AI agents.