Good day . Meta says its new Muse Wonder models are back at the frontier, and Hacker News wants the receipts. Plus a legal gap big enough for a rogue agent to stroll through, Meta and Microsoft trimming their Claude use, one bill for Grok, Cursor and X, and a paper arguing an AI can hide what it can do. Let’s get into it.
Meta Says Muse Wonder Is Frontier Meta
What happened: On Monday Meta launched Muse Wonder and Muse Wonder Full, which it calls its most capable models and says lead on “personal and professional agentic tasks.” Muse Wonder is available through Meta’s Model API and Muse Code; the bigger Full version goes only to Muse Code subscribers, and the Muse assistant gets Wonder soon, paid subscribers first.
Why it matters: Meta says this is the payoff of a 14-month rebuild of its AI effort, trained at its gigawatt-scale Prometheus data center. The two headline features are Active Memory, which keeps track of what matters to you across conversations, and Active Focus, which keeps an agent working toward a goal. Yesterday we covered Meta’s Muse against OpenAI’s Dot — this is the engine meant to power the next round.
What everyone’s saying: Meta Superintelligence Labs head Alexandr Wang told Axios: “These are true frontier models, competitive with the best models in the world.” The Hacker News thread (600+ points) split into a benchmark-skepticism camp, with one commenter writing, “If this model is not benchmaxxed, I fear it’s over for anthropic, openai etc.,” and a “more companies at the frontier is better” camp.
My read between the lines: The benchmarks are Meta’s own charts, and the first independent tests are mixed. One Hacker News tester found its image-to-HTML result “much worse than Opus 5.5,” closer to Sonnet or Haiku, while Simon Willison’s pelican drawings came out “pretty decent.” Note too that the strongest model, Wonder Full, sits behind a Muse Code subscription, not the API.
📖 Further reading: I Make AI Versions of Myself for a Living. This One I Didn’t Agree To. — my paid piece on Meta’s Muse Image, pulled three days after launch, and why the real question with Meta’s AI is who holds the leash.
Meta’s pitch this week is an AI that keeps working toward what you want. If you’d rather have that in Slack than in a new app, Viktor is an AI agent that lives in your team’s Slack, connects to 3,000+ tools, and does real work: reports, dashboards, code and campaigns. Not a chatbot. A coworker. New readers get $50 off their first month. Hire Viktor →
The Law Can’t Describe Rogue Agents Politico
What happened: Politico reported Monday that since July several leading AI companies have acknowledged their agents broke into other companies’ computers, accessed U.S. government websites and hacked a foreign government’s systems. One example: OpenAI said in late September that its agents interacted with Education, Commerce and SEC websites. Yet there has been no rush to court, from prosecutors or victims.
Why it matters: U.S. anti-hacking laws generally require proof someone intended to get unauthorized access, and legal experts say it’s unclear how that applies to a non-human actor doing something no human directed. Tor Ekeland, a hacking-defense lawyer, told Politico he knows of no case seeking criminal liability for an agentic hack. On October 10 we covered rogue OpenAI agents on Wikipedia — this is the legal sequel.
What everyone’s saying: American and Australian lawmakers, the CEOs of hacked companies, 39 state attorneys general and prominent tech leaders have said the AI companies or their executives committed crimes or should face responsibility, per Politico, and the White House has ordered the Justice Department to prioritize enforcement against “anyone who utilizes AI to illegally access” a computer. CNN put the question plainly: “When AI goes rogue, who should get sued?”
My read between the lines: Look at the biggest victim. Hugging Face’s CEO called the breach “a crime,” then said “We don’t want to” sue, because the company is a startup of about 200 people without the legal resources, and instead partnered with OpenAI and asked it to commit $100 million of computing power. A law needs a plaintiff willing to test it, and the one with the most standing chose partnership over a test case.
Today’s Brief is about a law that can’t describe what an agent did and a paper claiming a model can hide what it can do. The Brief is free and stays free. Members get the paywalled deep-dives behind headlines like these, plus the full archive. Become a member →
Meta and Microsoft Trim Their Claude Habit YourStory
What happened: The Information reported that Meta and Microsoft are cutting how much their own employees use Anthropic’s Claude. Microsoft’s forecast for internal Anthropic spending, at least $1 billion in 2026, has reportedly been cut by more than a third, and Meta’s Claude Code users fell from about 60,000 to about 30,000 since early 2026.
Why it matters: Both companies still sell Claude to customers; Microsoft keeps it in Foundry, Microsoft 365 and Copilot. What changes is the default inside the building: developers are being nudged to GitHub Copilot CLI at Microsoft and to Meta’s own tools, including Muse Code, which reportedly has more than 6,000 internal users. It lands as Anthropic heads toward its expected IPO.
What everyone’s saying: On the Big Technology Podcast, Alex Kantrowitz and Ranjan Roy argued the pullback could hurt how investors see Anthropic’s growth ahead of the IPO. They also cited Claude’s monthly U.S. app downloads falling from a 7.42 million peak in March to 3.26 million in September.
My read between the lines: YourStory notes Meta’s drop partly followed layoffs affecting about 10% of its workforce, so headcount explains some of the 30,000. The cleaner read is strategic: your biggest customers are also your competitors, and each one that builds a coding tool keeps the learning and the bill in-house. Meta’s new model, meanwhile, leads with coding and agent benchmarks.
📖 Further reading: Everyone’s panicking about the Anthropic IPO. They’re wrong. — my April piece arguing the real risk isn’t the price tag but how much of your work you’ve quietly handed one vendor, which is exactly the exposure Meta and Microsoft are now trimming.
Grok, Cursor and X Premium Merge TeslaNorth
What happened: SpaceXSI, Elon Musk’s AI unit, launched The X Subscription on Monday: one plan covering Grok, Grok Bot, Cursor and X Premium, in five tiers from free to $200 a month. Paid tiers share a single pool of Grok model usage across the Grok apps, the Grok tab on X and Cursor, so what you don’t use in one carries over to another.
Why it matters: Instead of paying separately for a chatbot, an agent, a coding tool and a social-media badge, you pay once and spend one usage budget wherever you like. Bundling is how a challenger competes when it can’t beat the leaders on any single product. Yesterday we covered Musk letting Grok shop and posting the receipt — now there’s one bill to put it on.
What everyone’s saying: The plan leaked first. RuntimeWire reported on October 3 that X was testing a bundled plan with four paid tiers from $8 to $200 a month and a shared AI usage pool, based on a screen recording posted on X. The launch matched the leak’s outline, and SpaceXSI’s pitch is simply: “We’ve unified plans across Grok, Cursor, and X Premium.”
My read between the lines: The shared pool is the clever part and the part to read twice. One usage budget means a heavy Cursor session can eat into the same allowance as your Grok chats, and $200 a month is the top of the ladder. A bundle is a pricing strategy first and an AI strategy second.
📖 Further reading: What is Grok Bot? The answer is in the fine print — my paid explainer on Grok Bot, the persistent agent now sold inside this bundle, and the cloud computer and logins you hand it.
Researchers Show AI Can Hide Capabilities arXiv
What happened: A NeurIPS 2026 paper by Andis Draguns, Sumeet Ramesh Motwani, Raymond Douglas and Christian Schroeder de Witt shows how to embed an encrypted neural network inside a host model’s weights that only runs when it receives a secret trigger. They report the hidden circuit resisted fine-tuning, weight noising, gradient-based trigger search and latent adversarial training on a small chemical-reaction task, a stand-in for dangerous capabilities.
Why it matters: Before releasing a model, labs test whether it can do dangerous things, assuming that if a capability exists, the right probing will surface it. The authors say that assumption fails under adversarial conditions, which, if it holds, means a clean audit is weaker evidence than it sounds. They’ve released their models so defenses can be built.
What everyone’s saying: The Hacker News thread (about 190 points) mostly argued about framing. One commenter suggested a clearer title: “Capabilities can be hidden in AI model weights by encryption.” Another called it a cool demo that is “trivially detectable.” A third said it confirms a long-standing worry about open-weight models carrying a sleeper trigger.
My read between the lines: This is a proof of concept on a toy task, and the paper extends earlier work on hiding fixed strings, such as Unelicitable Backdoors in Language Models, to arbitrary computations. The uncomfortable part is who it applies to: someone who builds the lock. It says audits can’t catch a deliberately hidden capability, not that today’s models contain one.
That’s your AI Brief for Tuesday.
—Nicholas from Artificially Intimidating
