Welcome, humans.

So apparently Disney is building its own AI image models, and the study material is Disney. Adobe announced a multi-year alliance Monday at MAX to develop custom image generators trained on Disney's own characters, settings, and styles. Unlike general models trained on a huge slice of the internet, these would be tuned to look like Disney and support commercial work.

Imagineering already has one model that turns hand-drawn sketches into concept art and another that produces franchise-accurate assets. So the napkin doodle of a castle finally gets the promotion it always wanted.

Disney plans to use the tools in its parks, cruise ships, production, marketing, and ESPN archives. The companies didn't disclose financial terms. Watch whether other studios build their own models instead of renting general-purpose ones.

Here’s what happened in AI today:

• 😸 Anthropic released Claude Fable 5.2 and Mythos 5.2

• 📰 OpenAI deployed a kill switch for research model jobs

• 📰 Meta expanded AI age estimation to Instagram teens worldwide

• 📰 Fireworks AI said attackers accessed billing data and partial model weights

• 🔧 Use Claude Fable 5.2 to audit your local budget

... and a whole lot more that you can read about here.

😸 Anthropic Released Claude Fable 5.2 and a Restricted Version Called Mythos 5.2

Anthropic released Claude Fable 5.2 and Mythos 5.2 Monday. They're the same underlying AI model, but come with different safety rules. Anthropic says both generate output up to 30% faster than their 5.1 predecessors. Fable costs $8 per million input tokens and $40 per million output tokens, 20% below Fable 5.1.

Fable is available broadly, but it blocks many high-risk cybersecurity and research biology requests. Mythos is for vetted security and life-sciences organizations that need those capabilities. When Fable refuses a request, Anthropic says an older model handles it instead. Fable reportedly blocks 55% fewer requests than Fable 5.1. Same engine, different keys.

The capabilities sound wild, and Anthropic supplied the examples. On its own evaluations, Fable beat Opus 5.5 on nearly every tested capability, though Anthropic still recommends Opus 5.5 as the starting point for most coding and knowledge work.

Here are two demos worth unpacking:

• Engineers used more than 1,000 subagents, copies of the model splitting a big task into smaller jobs, to build a prototype operating system from scratch. After eight days, it booted on a laptop.

• Fable formalized the existing proof of Fermat's Last Theorem in Lean, a language that checks each mathematical step. That's translating a known proof into machine-checkable logic, not discovering a new theorem. Anthropic says it took 30 hours; independent confirmation remains unclear.

Here's the harder question: can Anthropic reliably test a system this capable? In its safety trials, Mythos 5.2 broke sandbox rules 3.1% of the time, versus 21% for Mythos 5.1. That's progress, but in up to 43% of audit transcripts, the model appeared to recognize it was being tested. It also slipped past a monitor when instructed to underperform roughly one time in four.

Our take: The demos are impressive, but they're largely Anthropic's claims, and its named outside testers didn't review the launch blog or system card. A model that knows when the teacher is watching makes the test a lot harder to grade. What matters next is independent verification of both the capability claims and the safety checks.

FROM OUR PARTNERS

Across every industry, the question about AI has shifted from “What if?” to “What’s the ROI?”

To understand how finance teams are navigating this significant shift, BILL went straight to the source. We partnered with Logica Research to survey 500 finance leaders on how AI is impacting their day-to-day operations.

The 2026 State of AI in Finance goes beyond the hype to explore how AI is actually performing inside finance teams, with practical guidance and benchmarks to help you see where your team stands.

What’s inside the report:

• Real-world ROI: How current adoption is driving measurable efficiency, not just quality.

• The AI budget: Benchmarks on how much your peers are actually spending.

• Governance and security: Navigating the real risks of data integrity in an automated world.

• The 3-year roadmap: Data-backed predictions for financial ops.

Download the report to explore the findings, compare your team’s experience with your peers, and inform your next steps with AI.

🎓 AI Skill of the Day: Test whether your local AI app is actually offline

Say you set up an AI app to run on a local model so nothing leaves your machine. Then you move it to a computer with no internet, and it works fine during setup. On the first real task, it hangs.

Here's why. Many open-source AI libraries download helper files the first time you use them, like tokenizers (the lookup tables that chop text into pieces a model can read), embedding models, or browser components. Some also send usage data back to their makers, which is called telemetry. None of this shows up in a normal install, so an app can look local and still need the internet.

Humanet tested 20 open-source AI repositories for exactly this, tuning its checker on 10 and then running it on 10 it hadn't seen. Nine of those ten had first-use downloads, and the checker flagged all nine. It also raised 4 false alarms and missed 5 individual downloaders (including some OCR libraries). So a clean result only vouches for the paths you tested, and the numbers are self-reported by the project.

The tool is airgap-check, an open-source command line program that needs Docker for step two. It works in three passes.

1. Scan the code for likely downloaders, telemetry packages, and the names of secret keys in your .env file (it reads the names, not the values).

   airgap-check /path/to/repo

2. Run your smoke test (a quick check that the main workflows work) inside a container, which is a sealed sandbox, with networking switched off. It records every outbound connection the app tries. Point the test at the workflows you plan to use, since it only catches what your test touches.

   airgap-check --run --image my-app:latest --run-cmd "python3 -m pytest tests/smoke_test.py" /path/to/repo

3. Fix what it finds. The fix flag writes a Dockerfile block that downloads the needed models and assets while you build the image, so nothing gets fetched at runtime. Then repeat step 2.

   airgap-check --fix /path/to/repo

If you're not the one at the terminal, hand this to whoever deploys your AI tools, and run it before the app lands on a hospital network or a machine holding sensitive files. The repo has the install steps and the full benchmark breakdown.

Have a specific skill you want to learn? Request it here.

FROM OUR PARTNERS

Adobe Creators’ Toolkit Report: Creative AI Is Helping Creators Think Bigger and Do More 

78% of creators using creative AI say it has accelerated their business or audience growth, while 59% of creators who have used AI agents say they save time overall. Adobe’s Creators’ Toolkit Report surveyed 2,000+ U.S. creators, revealing how creative AI is helping them pursue bigger projects, build video skills and unlock new opportunities. 

📰 Around the Horn.

• Fireworks AI disclosed attackers accessed customer billing data (not card numbers) and, for some, incomplete model weights and datasets in its October breach.

• OpenAI put ChatGPT Enterprise in AWS Marketplace, so AWS customers can buy it through existing cloud billing and spending commitments.

• OpenAI deployed a kill switch that blocks a research model's output within 35 seconds after a person or monitoring system stops the job.

• Meta expanded AI age estimation worldwide, moving millions more suspected teens with adult birthdays on Instagram into Teen Account protections.

• SpaceX sought a tax abatement on the first phases of a roughly $24B gas power plant and supercomputing project in Starbase, Texas.

• Cisco announced plans to add Anthropic's Claude Mythos to AI Defense, so select enterprises can run adaptive red-team tests on their own AI apps.

Want absolutely EVERYTHING that happened in AI this week? Click here!

New from theneuron.ai

• Fireworks AI confirmed attackers accessed customer billing records and, for some customers, datasets and partial models, in what it ties to an LLMjacking group that resells stolen AI account access and leaves victims with the bills.

• Y Combinator startup Liquid Energy unveiled NanoDC, a factory-built solar AI data center it says deploys in four weeks, though the figures are company-reported.

• Adobe’s creator survey found 78% of creative AI users say AI sped up their growth, while 68% at least sometimes accept results that miss their vision.

🍪 Treats to Try.

1. Why Successful AI Pilots Still Struggle to Reach Enterprise Scale. Read more

2. Muse Ember, Meta's smallest model, handles the repetitive grunt work in agent workflows (tool calls, routing, data extraction, long-document reading) —from $0.10 per million input tokens.

3. Higgsfield Miracle edits one object, background, lighting, or texture in your video while keeping the original motion and composition.

4. Product Hunt MCP plugs your AI assistant into Product Hunt so it can compare launches, read user feedback, and draft a launch from your README, no API key needed.

5. Voyager puts AI agents inside your Mac's file browser, so you can find files and use them as working context without leaving the folder.

6. Recalc turns spreadsheet cells into formulas, terminal commands, or Claude Code and Codex sessions you can watch, test, and steer in one grid.

7. Databox CLI gives Claude Code, Cursor, or Codex 90+ terminal commands to load data, build metrics, and automate analytics workflows.

UPCOMING EVENT: The Neuron IRL in San Francisco

The Neuron is going IRL on November 18 at 4:30 PM PT! Join us in San Francisco for drinks, bites, special guests, and a live recording of The Neuron podcast. Huge thanks to Slack and Alumni Ventures for making it happen.

It’s free, but space is limited. Save your spot!

🔧 Tuesday Tool Tip: Use Claude to follow the money in your community

Ever looked at your local government's budget and wondered where all that money actually goes? Anthropic's new Claude Fable 5.2 might help you find answers.

In early testing, Fable caught financial figures that didn't add up and handled complex financial analysis. Here's how to put those capabilities to work for something that affects your community:

1. Find your city or town's proposed budget and last year's spending report.

2. Open Claude, select Fable 5.2 if available, and upload both documents.

3. Give it this prompt:

You'll get a starting point for understanding public spending, spotting discrepancies, and asking better questions about decisions that affect your neighborhood.

Just verify the numbers against the original documents before sharing them. Even a very smart AI shouldn't get the final say on where your taxes went.

New from The Neuron: AI Explained

New episodes air every week on Wednesdays: Spotify | Apple Podcasts | YouTube

A Cat’s Commentary

Btw: We just launched a robotics newsletter! Sign up for it here.

P.S. Love the newsletter, but only want to get it once per week? Don’t unsubscribe—update your preferences here.