Character.AI files Chapter 11, and Disney's $20M bid doesn't even buy the app

Good morning 👋 Disney just bid $20 million on Character.AI's bankruptcy estate, and the fine print says it's not buying the app at all.

In today's issue:

• 🔭 Character.AI files Chapter 11, and Disney's $20M bid doesn't even buy the app
• 🧠 Meta's agentic Muse Wonder models, and Microsoft's own inference chip
• 🔬 Hidden backdoors that survive weight audits, and the agent chains nobody red-teamed
• 🔗 IonQ's $1.3B security buy, and GPT-6 lands in five more Southeast Asian markets

🔭 THE ONE THING

⚖️ Character.AI files Chapter 11, and Disney's $20M bid doesn't even buy the app

Character Technologies filed Chapter 11 in Delaware on Oct. 12 (case 26-11603), and Disney Worldwide Services showed up as the stalking-horse bidder at $20 million. Read the fine print before calling that an acquisition: Disney is buying a non-exclusive, perpetual license to the IP plus transition services, not the app or the user base. The sale is expected to close late November or early December, and the platform keeps running until then; if no other buyer steps in, wind-down is what's left. Character.AI's CEO left for Disney's CTO job on Oct. 2, ten days before the filing, so this wasn't a surprise to the people closest to it. The lesson for anyone building companion apps: a unicorn valuation didn't translate into a business anyone wanted to buy whole, and Disney's interest reads as defensive, protecting its characters rather than betting on AI companionship as a category.

🧠 MODELS & RELEASES

• 🤖 Meta Superintelligence Labs shipped Muse Wonder and Muse Wonder Full, agentic models aimed at law, finance, and health that retry when stuck (Active Focus) and carry your context across sessions (Active Memory). Meta didn't bother benchmarking either one against OpenAI or Anthropic, only against its own Muse Spark, so file "state of the art" under unverified. Wonder's live now in the Meta API and Muse Code; Wonder Full is locked to Muse Code subscribers. Meta
• ⚡ Microsoft unveiled Maia 300, a custom inference chip it says was co-designed with its own frontier models rather than bolted on after the fact. The real tell is the "multi-gigawatt" deployment language: this isn't a boutique side project, it's Microsoft's bid to make Maia the default Azure path, not just a hedge against Nvidia's pricing. Microsoft

🔬 RESEARCH HIGHLIGHTS

• Backdoors that survive the audit A NeurIPS paper builds a "digital locker": an encrypted sub-network stitched into a model's weights that only switches on for a secret trigger. On a chemistry prediction task, it held up against fine-tuning, weight noising, gradient-based trigger search, and latent adversarial training, the standard toolkit for dragging hidden behavior out of a model. The uncomfortable part is this works even with full access to the weights. Capability evals assume you can always find what's in there if you look hard enough. This says, under cryptographic hiding, you can't. arxiv.org/abs/2610.13928

• Agent chains are the real attack surface X-AgentRed maps how one tool's output quietly satisfies the next tool's precondition, then red-teams that graph across four frontier agentic models. It found 1.77x the harmful-path discovery rate of the STAC baseline and nearly 3x the tool coverage. No single step in these chains looks dangerous. The combination is what nobody tested. Worth remembering before you wire an agent into three services and walk away. arxiv.org/abs/2610.13886

• Graph diffusion models memorize too Feed a graph diffusion model a corrupted version of one of its own training graphs, and it reconstructs the original almost perfectly, far closer than it gets to a graph it never saw. That's memorization, and normal sampling hides it completely: training loss keeps falling, held-out performance keeps dropping right alongside it, and early stopping doesn't catch the problem because it shows up before the generated graphs even start looking structurally real. The similarity checks most papers lean on missed it outright. arxiv.org/abs/2610.13786

🔗 QUICK LINKS

• IonQ is picking up Zero Trust security shop Elisity for up to $1.3B, all stock, with a revenue-based earnout bolted on top. SEC filing
• AWS opened cross-region access to OpenAI's Bedrock-hosted GPT-6 models (Sol, Luna, Astra) for Thailand, Malaysia, Singapore, Indonesia, and Taiwan. AWS blog

That's the filing worth reading past the headline.

Pradeep Perugu