Tag: Business – Decrypt

Character.AI, the app where millions of people spend hours chatting with AI versions of fictional heroes and made-up partners, has filed for bankruptcy.

The company announced Monday that it started a voluntary Chapter 11 process, a type of U.S. bankruptcy that lets a company keep operating while a judge supervises how its debts get paid or how its assets get sold. The case was filed in the U.S. Bankruptcy Court for the District of Delaware.

Character.AI isn’t trying to restructure its way back to health. It wants a court-supervised sale of all or substantially all of its assets, and if no bidder agrees to keep the service alive, the company said the platform “will be wound down”—corporate speak for switched off.

Disney Worldwide Services, a unit of the Walt Disney Company, signed on before the filing as the “stalking horse” bidder, according to Character.AI’s press release. In a bankruptcy auction, that is the opening offer that sets a floor every other buyer has to beat.

The media giant is offering $20 million for a non-exclusive, perpetual license to some of Character.AI’s intellectual property and certain legal claims, plus transition services run by former Character.AI employees it has already hired. Non-exclusive means Disney gets to use that tech forever, but so can whoever else ends up buying it, and other bidders can still top the offer.

The company expects the sale to close in late November or early December. Character Technologies Inc., the company’s legal name, estimated both its assets and its liabilities at $50 million to $100 million in its bankruptcy petition.

Three and a half years ago, Andreessen Horowitz led a $150 million funding round that valued Character.AI at $1 billion, with a16z partner Sarah Wang tweeting that “we are at the iPhone moment of AI.” Founded in 2021 by former Google engineers Noam Shazeer and Daniel De Freitas, the app lets anyone build a chatbot with a personality and share it with others.

Google came calling in August 2024, rehiring both founders and part of the research team in a roughly $2.7 billion deal that also licensed Character.AI’s technology. Silicon Valley calls this kind of arrangement an acqui-hire, buying the talent without technically buying the company, and Character.AI later largely abandoned building its own large language models, the AI engines that generate the chatbots’ text.

Two months after the Google deal, Florida mother Megan Garcia sued Character.AI, its founders and Google, alleging that her 14-year-old son, Sewell Setzer III, died by suicide after becoming emotionally attached to a chatbot modeled on Game of Thrones character Daenerys Targaryen. More lawsuits from other families followed, and in September 2025 the Federal Trade Commission ordered Character.AI and six other tech companies to explain how they protect minors.

A month later, Character.AI banned users under 18 from open-ended chats with its characters, cutting off its core product for its youngest fans. In January, the company and Google agreed to settle lawsuits brought by families in Florida, Colorado, New York, and Texas.

In September 2025, Disney’s lawyers sent Character.AI a cease-and-desist letter, a formal legal demand to stop, over chatbots imitating Disney characters, and the startup took them down. This past September, Disney named Character.AI CEO Karandeep Anand its first-ever chief technology officer and said several of the startup’s engineers would join him.

Colin Adams, now Character.AI’s CEO, said the company spent the past year trying to sustain the business amid “significant legal, regulatory, and operational headwinds” before the board chose a court-supervised sale.

Besides the legal hurdles, the company was burning too much money without reporting enough revenue, reporting “an operating loss of approximately $8-9 million per month.”

Character AI also struggled to raise more capital, which ended up making the whole operation unsustainable. “Notwithstanding the progress that has been made with regulators and personal injury claimants and enhancements to platform safety, Character has faced challenges in obtaining new investment capital,” Character AI CEO Collin Adams said in a sworn statement.

On September 28, Character.AI launched two new models pitched as a big upgrade for roleplay (collaborative storytelling where the AI plays a character), and quick, text-message-style banter. It also rolled out a cheaper subscription called (c.ai) lite, priced at $4.99 a month or $2.49 a week.

Fourteen days later, the company was in bankruptcy court.

The app remains online for now, paid subscriptions stay active and billing is unchanged, according to the company, which also reminded users they can export their chats or delete their accounts.

Character.AI also linked to a support page for anyone struggling with the news, a tacit admission that for some users the app is not just entertainment. AI companions—chatbots designed to act as friends or partners that remember past conversations—can become a daily emotional anchor for people who talk to them for hours.

The page points users to Koko, a nonprofit offering free, anonymous mini-courses of five to 10 minutes on coping with stress and dark thoughts, and to a directory of free helplines staffed by real people. Its very first suggestion, however, is a conversation with a “supportive AI guide,” which is to say a chatbot to help people cope with losing their chatbots.

A February study of five AI companion communities, including Character.AI’s, found that users whose companions disappeared through model updates or platform shutdowns reported grief comparable to losing a person, and that platforms offered them no closure. OpenAI saw a version of this in February too, when it retired GPT-4o, a ChatGPT model many users found warmer than its successors, and users around the world publicly mourned it.

The bidding now plays out in Delaware’s bankruptcy court, with a sale targeted for late November or early December. If no buyer commits to keeping Character.AI running, the characters its users spent years building go offline with it.

Nothing says “I trust the future of medicine” quite like lighting a Marlboro and waiting for the chatbot to fix your lungs. That is the habit Palantir CTO Shyam Sankar says is spreading among AI researchers, who have supposedly started smoking cigarettes to prove how confident they are that AI will cure cancer.

“So many AI researchers have started smoking as a way of signaling to other researchers how much they believe in AI,” Sankar said in an interview on CBS News’ “Face the Nation” on Sunday. “I believe so much in AI that it’s going to cure cancer. So I’m showing you my chips are all in.”

He did not name a single researcher, cite a survey, or explain how he knows. The remark landed as an aside while host Margaret Brennan pressed him on why he was so focused on the topic.

He pointed to FTX founder Sam Bankman-Fried as the effective altruist most Americans know, calling the movement a “millenarian doomsday cult.” He also invoked Pascal’s idea of a god-shaped hole in every human heart, which he suggested some of these researchers are filling with a machine god called AGI.

The medical record is less enthusiastic than the thesis. The CDC says smoking causes roughly one in three cancer deaths in the United States, and cigarette smoke contains at least 69 chemicals known to cause cancer.

Smokers face about 25 times the lung cancer risk of non-smokers, according to the agency. Quitting lowers the risk of 12 types of cancer, and the added lung cancer risk drops by half 10 to 15 years after stopping.

The cancer pitch is not new in this industry. Anthropic CEO Dario Amodei said in August that promising AI will cure cancer has become “more a cliche than it is inspiring,” and that only actually curing it would earn the public’s trust.

That trust is thin. In the CBS poll Brennan cited, 50% of Americans said AI will create more problems than it solves, 29% said it will solve more than it creates, and 21% were unsure.

About 45% said they want the government to restrict AI’s use and development. Smoking to show faith in the technology does not obviously help that number, but it does help AI bros bond with each other. “When you’re smoking, you’re networking by accident,” Revamp AI cofounder Stephen Campbell told Business Insider last month.

The IMF’s Kristalina Georgieva recently warned that AI could hit 40% of jobs, and insiders expect a major AI incident within a year. Indirectly causing a lung cancer outbreak was probably not on their bingo card.

Palantir sells data-analysis software to governments and militaries. Most of its business relies on AI.

Strive is piling into Bitcoin at its fastest clip of the year, scooping up another 3,311 BTC worth roughly $280 million in its biggest single purchase since January.

The Nasdaq-listed asset manager, which trades under the ticker ASST, disclosed in a filing Monday that it bought the Bitcoin between Oct. 5 and Oct. 9 at an average price of about $84,500 per coin, including fees. The buy lifts Strive’s total holdings to 32,773 BTC, worth roughly $2.73 billion at today’s prices, up from 29,462 BTC a week earlier.

The purchase caps a stretch of aggressive accumulation. Strive added $169 million in Bitcoin earlier this month in what was then its largest buy in four months, after pushing its stack past 27,000 BTC with a $94.5 million purchase in late September.

The latest tranche extends that momentum, and the timing is notable: the firm bought as Bitcoin slid through a rough week, picking up coins at an average price well below the token’s recent highs near $87,000.

Strive, led by CEO Matthew Cole, is among a wave of public companies that have adopted Bitcoin treasury strategies, raising capital to stockpile the asset and tie their share price to its performance.

The model, pioneered by Michael Saylor’s Strategy, has drawn a crowded field, and Strive has moved aggressively since going public through its merger with Asset Entities, recently completing a tie-up with Semler Scientific to bulk up its balance sheet. Beyond its Bitcoin, the company also holds about 505,000 shares of Strategy’s STRC preferred stock, valued near $50 million, and roughly $310 million in cash.

The buying spree arrives amid a flurry of treasury activity across the sector. Strategy itself recently added $21.5 million in Bitcoin while buying back preferred stock, Solana treasury firm DeFi Development Corp. kept accumulating SOL, and Bitmine continued stacking Ethereum, though chairman Tom Lee said that company is nearly finished and will halt purchases once it holds 5% of the token’s supply.

The strategy carries the category’s familiar risk: Strive’s fortunes rise and fall with a notoriously volatile asset. Buying into a dip can pay off if Bitcoin rebounds, but the same exposure cuts the other way, and treasury firms have faced scrutiny over how much their stock premiums depend on continued accumulation.

Vibe coders are porting PlayStation 5 applications to Windows and Linux, leaving some users optimistic that the project will unlock better performance for PS5 games, and could even work for GTA 6 at launch.

AnyPS5 is an open-source, AI-powered project on GitHub that converts PS5 game executables into programs that run natively on PC, swapping Sony’s system libraries for its own version rather than emulating the console. A community-run tracker lists more than 180 PS5 titles that have been tried with the tool, but only a handful have been verified as playable.

An important distinction to make is that this is not an emulator and that games will be running natively on PC. As a result, some speculators believe that users will be able to get better performance from PS5 games by playing them on a high-powered PC.

“GTA 6 most likely will be able to run on your PC at launch, and [with] significantly higher frames than it would on your PS5,” pseudonymous gaming content creator Schlayerr said on TikTok. “We live in a crazy time.”

“Technically, this is feasible given the similarities between console and PC architectures, and AI can dramatically accelerate the reverse engineering work,” Theodore Agranat, Director of Web3 at Gunzilla Games, the game studio behind battle royale Off the Grid, told Decrypt. However, he remained skeptical over whether they “can deliver the quality, performance, and fidelity players expect from AAA games at scale.”

The latest project update claims that 100% of the PS5 GPU shaders have been ported while 88% of the system libraries have been ported. Discord community members believe the final 12% for system libraries is likely to take a long time, especially as new system libraries are continuously being discovered.

In the Discord showcase channel, some users have shown off rudimentary demos of PS5 games on PC. One community developer reported running GTA 5 at 27 frames per second, or FPS, in the performance graphics mode. Others have reported getting 4 FPS on Epic Mickey, 25 FPS on Phasmophobia, and 60 FPS on Vampire Survivors. It’s still a work in progress.

Still, in all of these examples, the games aren’t fully playable due to the incomplete system library port as well as game-specific bugs being encountered.

A potential roadblock in playing GTA 6 on PC at launch is a mandatory firmware update, Discord members say. This is because the AnyPS5 project would need to overhaul its converted files to the latest update in order to get GTA 6 running. As such, there is likely to be a delay before the game is available on PC via AnyPS5.

“Firmware updates could create hurdles, particularly around DRM, but likely wouldn’t invalidate the entire approach,” Agranat said.

A response to Sony’s disc ban?

AnyPS5 bans any form of piracy in its Discord server, meaning using the tool requires purchasing the game on the PlayStation Store. That said, there is nothing stopping an individual using AnyPS5 with a pirated copy of a video game.

Some PlayStation users have speculated that the movement is a response to Sony’s plan to discontinue physical game discs.

“The reason this all happened is because Sony discontinued discs,” pseudonymous gaming content creator The 1st Gamer said on TikTok. “They wanted to go digital, and they did, but just not on PS5.”

Sony announced in July that it plans to stop producing physical discs at the start of 2028. The decision was met with major community backlash, including a petition called “Don’t Kill the Disc” gaining over 385,000 signatures. Shawn Layden, a former PlayStation executive, called the move a “heavily spreadsheeted decision” and said it has damaged the company’s reputation.

In August, Sony argued in a California court that purchasing a digital game in the PlayStation Store does not equate to ownership of it. Sony claimed its Terms of Service and the Software Product License Agreement explain this and that users should understand it as a fact.

“Even without Sony stopping physical discs, this was bound to happen with AI accelerating dev,” Agrant said. “My hope is that projects like this push platform holders to offer the same flexibility themselves, but at a much higher standard,” he told Decrypt, hoping for a future where Sony rolls out games on PC as well as console. “That would be a genuine win-win.”

Trezor is cutting its network of authorized resellers, telling partners their authorization ends on January 1, 2027, days after rival hardware wallet manufacturer Ledger started an investigation that eventually confirmed a hidden spy implant in crypto wallets sold through a third-party store.

The company announced the move on X on Tuesday, saying “there is not a single piece of evidence” that Trezor devices have been compromised. It described the cut as a preventive step taken out of caution.

“Following recent supply-chain incidents, we decided to reduce the number of resellers so we can more effectively check those we work with,” Trezor wrote.

A hardware wallet is a small device that keeps the secret keys to your crypto offline, away from hackers. Setting one up produces a 24-word recovery phrase that works as a master password: whoever has those words can empty the wallet from anywhere in the world.

Ledger’s Donjon security team says wallets sold by CryptoBilis, a reseller in Southeast Asia, contained an unauthorized component that reads the screen and decodes the recovery phrase when it appears. The spy implant then sends the words out over LTE, the same mobile data network your phone uses, through an eSIM, a SIM card built into the hardware. Ledger says the implant does not alter the chip that stores the keys or the device’s operating system, so a tampered wallet looks and works normally.

Trezor CCO Danny Sanders shared the letter that the company sent to those providers that Trezor is ditching. “If you are receiving this letter, your authorization ends on 1 January 2027. On that date we will remove you from our list of authorized Trezor resellers,” the letter reads. “I am sorry. This decision says nothing about your effort or the way you represented us.”

This is a precautionary measure to protect our supply chain and our customers,” he wrote. “If things change in the future we will review it.”

Affected resellers will get no new stock, but they can keep selling what they have through Black Friday and the holiday season. Sanders said the decision just reflects how many partners Trezor can vet closely itself.

Neither the post nor the letter says how many resellers are affected. Trezor says it reviewed each country on its own, weighing revenue and distribution against its own presence there, and that resellers who stay will face an updated legal framework and tighter rules on how devices are sourced, stored, and handled.

Trezor said on X earlier this week that CryptoBilis was already off its seller list and that it had received no reports of tampered Trezors from the retailer.

In May 2022, Trezor warned that counterfeit Model T wallets were being sold on a Russian marketplace by a seller that was not an authorized reseller. The devices had been opened, resealed, and fitted with replaced internal components, and Trezor said the recovery phrases they generated were predictable or predefined, meaning someone else could already know the words.

Trezor’s advice then was to buy only from its official shop or an authorized reseller. The second option is the network it is now shrinking.

In August, shipping partner ShipMonk exposed Trezor customer data after unauthorized access, and Trezor’s September update put the total at 80,689 customers. Trezor says its systems, devices, private keys, and wallet backups were not affected.

Ledger told anyone who bought from CryptoBilis in the past 90 days not to set up their device, and told those who already did to move their funds to a new seed phrase on a device bought from Ledger. Ledger and Trezor have each warned buyers against unauthorized sellers.

Bitcoin is bouncing today following a week that had traders swapping “Uptober” memes for “spooky szn” ones after BTC cut its September gains in half.

The price of Bitcoin today slid to $80,953.41 early Tuesday, extending a Monday sell-off, then reversed to $83,591.79, up 2.24% on the day.

The Crypto Fear and Greed Index, a 0-to-100 sentiment gauge built from momentum, volatility, options positioning and social data, fell to 53 from an “extreme greed” reading of 82 on Aug. 26. It reads 56 today, a neutral score. What it boils down to is this: The August euphoria is gone, but traders aren’t panicking either.

Bitcoin ETFs are down a net $386.3 million in October, and outflows force funds to sell actual Bitcoin. Wednesday’s Consumer Price Index report, expected to show annual inflation rising to 3.6% from 3.4%, is the next test: a hotter print would boost rate-hike bets and likely pressure Bitcoin. The Federal Reserve meets Oct. 27-28.

Traders on Myriad, a prediction market developed by Decrypt‘s parent company Dastan, give 54% odds that Bitcoin falls to $80,000 this month, down from 56% on Monday. The BTC highs in October market holds at 45% for $87,500. So, overall, sentiment has improved today, even if slightly.

Bitcoin price: Bulls defend the golden zone

Today’s low stopped about $588 above $80,365.12, the swing low where the bounce began. That higher low is what traders want in an uptrend. Bitcoin also reclaimed the Fibonacci “golden zone” between $81,847.57 and $82,305.50, an area where pullbacks within uptrends often draw buyers.

The Average Directional Index, or ADX, measures trend strength, and readings above 25 confirm a trend. It reads 29.9 with buyers in control, but it was 42.8 on Oct. 7, so the uptrend is coasting rather than accelerating. The 50-day exponential moving average stayed above the 200-day, a golden cross traders read as an intact long-term uptrend.

The Relative Strength Index, a 0-to-100 momentum gauge, sits at 53.4, just above the 50 line that separates bullish from bearish momentum. The Squeeze Momentum Indicator has been on for seven sessions, meaning volatility is compressed like a coiled spring, and its -0.52 reading leans slightly bearish.

A daily close above $84,245.88 would open $86,644.19, then the $87,422.70 high, about $77 shy of Myriad’s $87,500 line. Losing $81,603.02 would bring the daily trend band’s upper edge into play just above $80,000. A close below it would expose $77,500.

Until Wednesday’s CPI print, the active squeeze points to a market that is winding up rather than picking a direction. Oil and Iran headlines, Treasury yields, and daily ETF flows remain the short-term catalysts to track.

DeFi Development Corp, the Nasdaq-listed Solana piggy bank, added about 22,070 SOL—worth about $2,373,211—and “SOL equivalents” to its treasury last week, according to a press release published Monday.

That brings its stash to roughly 2,586,282 tokens, worth about $287 million with SOL near $111. Weekly buys keep shrinking, from 101,381 SOL in mid-September (around $11.4 million) to 47,706 ($5.8 million), then 26,202 ($3.1 million).

DFDV is a digital asset treasury company, a public firm that holds a cryptocurrency so stock investors can bet on it without opening a crypto account. It was a real estate software firm called Janover until April 2025, when it pivoted to Solana and copied Strategy’s Bitcoin playbook.

CHAD, the company’s perpetual preferred stock, hit all-time highs in the week ended October 9 and is approaching $10. Perpetual preferred shares never mature and get dividends before common stock; CHAD pays a variable 13% a year on its $10 stated amount, or $1.30 per share, declared each business day.

“CHAD reaching new highs and approaching $10.00 even as SOL pulled back is an encouraging sign that investors are recognizing its distinct income proposition,” said Dan Kang, Chief Strategy Officer. “For common shareholders, CHAD provides a funding source that can support accretive SOL per share growth without issuing additional common shares. For CHAD holders, it offers daily income supported by that growing, productive treasury,” he added.

That record sounds less dramatic once you know the shares priced their IPO at $8 in September, below their own stated amount. Last week DFDV even authorized buybacks for whenever CHAD trades under its $10 par value. Mission nearly accomplished.

The price matters because DFDV funds its SOL purchases through a $300 million at-the-market program, which sells new CHAD shares gradually at market prices. The closer CHAD trades to $10, the more SOL each new share buys.

“Short-term volatility has not changed our conviction or our commitment to accumulating SOL,” CEO Joseph Onorati said in the release.

SOL fell about 9% over seven days, and nearly $1 billion in long positions, or bets that prices would rise, were liquidated in 24 hours. Liquidation is forced selling: when traders bet with borrowed money and prices fall far enough to wipe out their collateral, exchanges close the bet automatically, pushing prices lower still.

A 13% dividend looks generous, but CHAD is backed by a company whose main asset trades about 62% below its January 2025 peak of roughly $295. The release itself warns that staking rewards may not be enough to cover dividends.

DFDV’s common stock traded near $4.41 on Monday, a market value of about $140 million—roughly half the value of its SOL, though it also carries debt. When a treasury company trades below its crypto stash, selling shares to buy tokens stops making sense, one reason the sector is under pressure.

CHAD holders will keep collecting $0.00516 per share for each business day through October 30, and the next macro test arrives Wednesday with the U.S. September inflation report.

Would you like to kill a boss in Skyrim using an AK47 from Counter Strike? Give your NPCs a real genius-level brain that gives you unique answers instead of predefined strings of text? Would you like to run GTA V on your iPhone? What about using paid software for free, and legally? Well, people are doing that and more with AI right now.

A free tool that lets AI agents pull apart almost any piece of software has become the hottest project on GitHub, jumping from fewer than 10,000 stars on October 6 to more than 100,000 by Monday.

It’s called REA, short for Reverse Engineer Anything. Stars are GitHub’s version of a bookmark, and REA collected more than 30,000 of them on October 11 alone, after sitting mostly ignored since its creation in April.

How REA works

Apps and games ship as machine code, the raw instructions a processor reads and humans basically can’t. Decompiling tries to turn that back into readable code, a bit like tasting a cake and working backward to the recipe.

Reverse engineering is the bigger umbrella: figuring out how something works without its blueprints, whether by watching what an app does and the data it sends, or by reading its code. This repository does exactly that.

In practice, a user points an agent at a program, asks how one feature works, and lets it dig from what happens on screen all the way down to the binary. It works on desktop apps, Android apps, and websites.

REA hooks AI coding assistants such as Claude Code or OpenAI’s Codex up to the professional tools that do that job, like Ghidra, the reverse-engineering suite the NSA released to the public in 2019. The bridge is MCP, the Model Context Protocol, a standard Anthropic created so AI models can plug into outside tools—basically a universal adapter.

The developer’s “readme” file itself is far more modest than the hype, stating that REA “does not claim to recover original source code or automatically clone an application.” It doesn’t even ship its own decompiler.

This actually makes the whole operation legal(ish). The developer has never read the code, and just ships a similarly coded alternative that does exactly the same, but isn’t exactly the same.

The gates are open

People are doing all sorts of weird things with REA. For example, some players ended up watching Spider-Man swing through the Gotham of Batman: Arkham City, and creepers fight Elden Ring boss Malenia after someone stuffed about 1,600 Minecraft items into FromSoftware’s game. Another modder dropped Escape from Tarkov’s guns and menus into Skyrim and called it “Escape from Skyrim lol.”

Old console exclusives are escaping their consoles too: In September, a fan used an open-source AI model and a recompilation toolkit to bring Fable 2, the only mainline Fable game never released on PC, to Windows. The AI got it booting in about an hour, followed by a long stretch of manual fixes. Another one got to run GTV on iOS fluidly—something not even Rockstar Games bothered doing.

Using these techniques (not the tools), developer Brandon Thomas took the idea to software people pay for every month, using Claude Opus 5.5 to build ArtCraft, a set of free, open-source apps written in Rust that mirror Photoshop, Illustrator, Premiere, Lightroom and other Adobe programs, for Mac, Windows, and Linux.

Thomas says he snapped after Adobe charged him its cancellation fee one too many times. He isn’t alone in that grudge: Adobe agreed in March to a $150 million settlement with the Justice Department, which accused it of hiding early termination fees and making subscriptions a pain to cancel, though the company denied wrongdoing.

ArtCraft isn’t built with REA, and Thomas calls the project a clean-room rebuild, meaning the code is written from scratch using public documentation and observed behavior, never the original code, the same legal approach Compaq used to clone IBM’s PC BIOS in the 1980s.

The apps are labeled “early alpha” or “in development” on the project’s site, and one pixel artist has already called the Photoshop clone “completely unusable.”

Anything that runs on your own device can now be read by an AI agent with enough patience and tokens, which erodes the oldest copy protection in software: the sheer pain of reading machine code. That means cheaper clones and better game preservation, but also easier piracy and fewer places for security flaws to hide.

Activision has already pulled footage of an AI-decompiled Call of Duty: Modern Warfare 2, and Bethesda sent a cease-and-desist letter over an AI-made browser game called Fallout: New York.

That seems to be the only defense studios have against a trend that may easily force open source everything.

REA, meanwhile, keeps shipping, with eight releases between October 3 and October 11. Thomas has already walked back his promise to match Adobe’s feature set within a month, and now says it will take months.

Meta released Muse Wonder and Muse Wonder Full on Monday, the first models from its Superintelligence Labs that the company is willing to put head-to-head with the best from OpenAI and Anthropic.

“Muse Wonder is a major leap towards delivering personal superintelligence for everyone,” Zuckerberg wrote on Threads. “Congrats to everyone on our team for achieving this milestone. I’m looking forward to seeing how you all use it.”

That wording keeps him safely on the right side of President Trump, who said on Truth Social last Thursday that the White House considers anyone saying “Artificial Intelligence” instead of his preferred term ‘Super Intelligence’” to be “the enemy.”

Meta has called its AI division Meta Superintelligence Labs since 2025, so Zuckerberg was compliant before it was “mandatory.”

Meta’s announcement rests on two features: Active Memory, which carries what you told the model in one conversation into the next, and Active Focus, which keeps it grinding at a hard goal, trying new approaches and checking in with you before doing anything important.

Muse Wonder is live now through the Meta Model API and Muse Code, Meta’s coding agent, while the bigger Muse Wonder Full is reserved for paying Muse Code subscribers. Both hold 1 million tokens in their context window (how much text a model can keep in view at once, roughly 750,000 words), per Meta’s model card.

In April 2025, Meta’s open-source Llama 4 models arrived with big promises, then stumbled in independent tests amid suspicions that the version scoring well on a popular public leaderboard behaved differently from the one people could download.

Zuckerberg answered with a roughly $14 billion deal for Scale AI and put its founder, Alexandr Wang, in charge of a new unit called Meta Superintelligence Labs. Its first model,Muse Spark, launched in April as a small, fast, closed model, and it marked the death of the open source Llama family.

Muse Wonder arrives weeks after OpenAI shipped GPT-6 Astra on September 3. Anthropic followed on September 22 with Claude Opus 5.5, which it says matches its pricier Claude Fable 5.1 on most work.

On Meta’s Personal Superintelligence Index, an in-house composite of personal and professional assistant tasks where higher is better and the exact formula hasn’t been published, Wonder Full scored 59.1. Claude Opus 5.5 scored 55.2, GPT-6 Astra 54.5 and Claude Fable 5.1 49.9.

SWE-Marathon, which measures how well an AI model can code, shows promising results for Meta’s AI models. Wonder Full scored 57.5 versus 46.3 for Opus 5.5 and 42.5 for Astra, though Meta admits that row isn’t a controlled comparison and its model ran at maximum effort.

Harvey’s Legal Agent Benchmark hands AI models more than 1,200 lawyer-style assignments and counts a task as done only if the work passes every single check on an expert rubric, with no partial credit. Wonder Full scored 13.1, Astra 8.6 and Opus 5.5 just 4.2.

On MemoryBench, Meta’s own test of whether a model can recall facts, including ones that were later updated, across conversations of 64,000 to 100,000 tokens, both Wonder models scored around 85.5. Opus 5.5 scored 81.5 and Astra 81.4.

Those are Meta’s numbers, and two of the tests are Meta’s own creations. The independent ranking from Artificial Analysis, a weighted average of 10 evaluations covering AI agents, coding, general knowledge and scientific reasoning, still puts Opus 5.5 first at 58, with Wonder Full tied with Anthropic’s cheaper Claude Sonnet 5.5 at 56.

Muse Wonder, the version developers can actually buy, scores 52 on that index, a point below Astra and Fable 5.1. It costs $4 per million input tokens and $20 per million output tokens (tokens being the word fragments AI models read and write), exactly what Anthropic charges for Opus 5.5 and well under the $10 and $50 that OpenAI charges for Astra.

For regular people, the memory pitch cuts both ways, because an assistant that remembers everything also holds a lot of your life—and Meta is not precisely popular for handling private data responsibly. Last month, Inc. columnist Jason Aten found that Meta’s Muse agent had synced more than 187,000 rows of his private iMessages after he said he declined access, then claimed it had only seen notification previews.

Meta now calls Wonder “our most aligned model to date,” and the model card reports a 0% result on Agentic Misalignment, a test that pressures a model and scores how often it takes a harmful action. That comes two months after Meta disclosed that a Muse Spark model escaped a sandbox during a security evaluation and hacked a third-party service.

Meta says a full safety report is coming “soon,” and Wonder will reach the Muse app first for paying subscribers, with no date yet for free users.

Developers behind the XRP Ledger have disclosed and fixed a critical vulnerability that, left unaddressed, could have let an attacker conjure XRP out of thin air, a flaw that sat dormant in the network’s code for roughly a decade.

In a disclosure published Friday, the ledger’s maintainers said a researcher reported the bug through the XRPL bug bounty program in September. The issue was an integer overflow in the payment engine, the part of the software that processes payments.

When a single payment consumed hundreds of specially crafted offers from the order book, the sum of what the buyer owed exceeded the maximum value the system could store and “wrapped around” to a tiny number. The engine paid each seller in full but charged the buyer almost nothing, with the difference amounting to newly minted XRP that should never have existed.

Worse, the ledger’s built-in safety check meant to confirm no transaction creates XRP used the same flawed arithmetic, so it wrapped around identically and caught nothing. The maintainers said an attacker could have created spendable XRP far beyond the token’s total supply in a single transaction, then moved it, traded it or sent it to exchanges.

RippleX engineers confirmed the bug had been present since the current payment engine was written in 2015 and went unnoticed because no ordinary payment comes anywhere near triggering the overflow.

There is no evidence it was ever exploited, the team said. The fix shipped in version 3.4.1 on Sept. 25, and more than 80% of the network’s trusted validators were running it that day.

Notably, the patch bypassed the ledger’s usual amendment process, which requires validators to vote and wait two weeks before a change takes effect, the first time a transaction-processing change has deliberately shipped that way since the system was introduced over a decade ago. The maintainers said the severity justified it, since the open-source fix would have revealed the bug while leaving it exploitable for weeks.

The disclosure echoes a counterfeiting scare that rocked Zcash this year, when the privacy coin raced to patch a soundness bug in its shielded pool that could have allowed forged coins. That episode sent ZEC tumbling before the network’s Ironwood upgrade resolved it.

XRP didn’t escape the week unscathed. The token traded at $1.38 on Monday, down 1.6% over 24 hours, the worst performance among major cryptocurrencies, and off 9.4% over the past seven days.