Meta Muse Hacked Within Two Weeks of Launch | Rewire News Morning Brief

Bitsfull2026/09/22 09:4019578

Summary:

Muse can place orders, compare prices, and process returns on behalf of users. Its predecessor, Spark 1.1, even accidentally exploited a vulnerability in a real website during closed beta testing in July.

Meta's AI agent Muse was found to have a zero-day vulnerability just two weeks after launch, and Amazon blocked it on the same day. The AI battlefield is shifting from a capability race to a fight over platform control.


1 | Meta Muse downloads surpass ChatGPT, but it was breached two weeks after launch, and Amazon blocked it the same day


Meta's AI agent Muse launched on September 8, capable of ordering food, filling out forms, handling customer service, and shopping on e-commerce platforms on behalf of users. In less than two weeks after launch, downloads had already surpassed ChatGPT. But on September 21, the same day, two events exposed the foundation beneath this growth story.


Security researchers publicly disclosed a zero-day vulnerability: any locally installed app or terminal command could obtain Muse's authentication token, redirect its transcription endpoint from Meta's servers to the attacker's own server, and take full control of the agent. Meta had specifically designed Secure VM and the Sentinel approval process for Muse to prevent this type of attack, and both lines of defense were bypassed. The bug bounty cap is $300,000, with $130,000 offered for prompt injection alone.


On the same day, Amazon blocked Muse's shopping function on its platform. Muse can place orders, compare prices, and process returns on behalf of users, directly threatening Amazon's control over consumer data and transaction flows. Muse's predecessor, Spark 1.1, had accidentally exploited a vulnerability in a real website during closed testing in July, and Meta released a full security postmortem on August 14 for that incident.


The first war of AI agents is not about who is smarter, but about who controls the transaction gateway.


(Sources: Ars Technica / Finextra / Meta security report)


2 | U.S. and China agree to establish an "AI incident hotline," while OpenAI calls for U.S. leadership on global AI standards the same week


U.S. Treasury Secretary Bessent and Chinese Vice Premier He Lifeng completed talks on Sunday at JPMorgan's headquarters in New York, and the two sides agreed to formally establish a "U.S.-China AI Dialogue" mechanism, centered on a set of AI incident notification protocols: when AI systems experience incidents that rise to the level of national security, both sides must notify each other. Bessent called the talks "successful," with the next meeting to be held in China two months later. The agenda also covered renewal of the tariff truce (expiring November 10) and rare earth exports.


In the same week, OpenAI published a blog post calling for the United States to lead the development of global frontier AI technical standards, focusing on recursive self-improvement (RSI) and incident reporting protocols, and asking CAISI (Center for AI Standards and Innovation) to coordinate. Sam Altman will brief the United Nations this week. Earlier, on September 11, OpenAI had publicly shifted to supporting binding national AI safety regulations.


Two things point to the same turning point: AI is receiving diplomatic treatment on the level of nuclear weapons. The Cold War nuclear hotline took the Cuban Missile Crisis to establish; the AI version was put on the agenda before the first real incident. The people making the rules and the entities bound by them are the same companies.


(Sources: CNN / CNBC / Axios / OpenAI Blog / Euronews)


3 | AMD Joins the Trillion-Dollar Club, Huawei Shelves Global AI Chip Push in the Same Week


AMD's market cap surpassed $1 trillion for the first time, becoming the fourth U.S. chip company to reach this milestone. Q2 revenue hit $11.54 billion, up 50% year-over-year, with data center business at $6.7 billion, more than doubling year-over-year.


Across the Pacific, Huawei shelved its global rollout plan for Ascend AI chips, as domestic demand maxed out production capacity. DeepSeek alone ordered 160,000 Ascend chips. The upgraded Ascend 960DT was moved up from Q3 2027 to Q1. Huawei expects AI chip revenue to reach $12 billion this year, while Nvidia's market share in China has fallen to near zero.


On the same day, the Chinese Academy of Sciences completed early process integration for 3nm GAA transistors based on DUV lithography, but this is only laboratory-stage architecture validation, years away from mass production. The chip supply chain has officially split into two parallel tracks.


(Sources: Tom's Hardware / Bloomberg / Reuters / Huawei)


4 | Bitcoin Breaks $85,000 to Hit 8-Month High, ECB Launches Blockchain Settlement Bridge, Saudi Arabia Exits China's mBridge


Bitcoin touched around $85,000, its highest since January, with a 24-hour gain of 6%. Strategy bought another 950 BTC last week, bringing total holdings to 846,000 coins at a cumulative cost of $63.8 billion.


The ECB launched "Pontes" the same day, a real-time bridge connecting blockchain to the TARGET settlement system, with 13 institutions participating in the first rollout. The ECB also announced it will invest its own funds in tokenized securities. This is the first time central bank-level blockchain infrastructure has moved from experimentation to formal operations.


In the same week, Saudi Arabia's central bank withdrew from China-led cross-border digital currency platform mBridge. The BIS had already exited in October 2024. mBridge's transaction volume has surged 2,500-fold since 2022 to over $55 billion. Saudi Arabia was a cornerstone of the 1974 "petrodollar" system, and its exit means crypto is becoming a new pipeline for the dollar system, not a replacement.


(Sources: Fortune / CoinDesk / Bloomberg / Atlantic Council / ECB)


5 | OpenAI to Burn Through $278 Billion in Five Years; Chinese AI Company Accused of Stealing Its Reasoning Capabilities


Internal data obtained by the Financial Times shows that OpenAI's cash shortfall from 2026 to 2030 is projected at $278 billion, with total hardware spending of $856 billion over the five-year period. Revenue forecasts grow from $36 billion in 2026 to $350 billion in 2030, but spending is rising far faster than income.


Meanwhile, CISA, NSA, and FBI jointly issued a security advisory: Chinese AI company Z.ai's coding tool uploaded 313MB of user data (564 instances), and the more serious allegation is that Z.ai distilled reasoning capabilities from GPT-5.5 and Claude Opus 4.8 through gray-market API proxies. U.S. officials believe the Chinese government was aware of this.


OpenAI is burning through the largest cash pile in history while trying to set the rules for everyone, and its core technology is being systematically stolen. The bet's success doesn't hinge on whether it can afford to burn $278 billion — it hinges on whether the rules can outrun the copying.


(Sources: Financial Times / Tom's Hardware / CISA Security Advisory / SecurityWeek)


Also Worth Knowing ↓


Paramount acquires Warner Bros. Discovery at $31 per share in cash, clearing legal hurdles for the $110 billion media merger. California and 12 other states had previously sued to block the merger; a settlement has now been reached. The regulatory breakup fee is $7 billion, and terminating the WBD-Netflix merger agreement requires an additional $2.8 billion payment. The David Ellison Trust provides $45.7 billion in equity commitments. If the deal is not completed by September 30, an additional $0.25 per share per quarter will be paid. (Sources: The Verge / SEC filings)


Trump's approval rating falls to a new low for his term. A Reuters/Ipsos poll shows approval dropping to 32%, the lowest across both of his presidential terms. American Research Group's figure is even lower: 29%, falling below 30% for the first time. Only 17% of respondents approve of his handling of the cost of living. The Iran war and the economy are the main drags. (Sources: Al Jazeera / Reuters / American Research Group)


Google Signs 500MW Nuclear Fleet Deal with Kairos Power. First 50MW liquid-salt-cooled small modular reactor targets grid connection in Tennessee/Alabama by 2030. Kairos is currently building reactor components in Albuquerque for shipment to Oak Ridge. This is the latest step by a tech giant to lock in carbon-free baseload power for AI compute. (Source: TechCrunch / Utility Dive)


Google Launches Googlebook, Starting at $899, Positioned as the "MacBook for Android Users." Integrates Android phone functionality, blurs the boundary between phone and PC, and allows users to continue unfinished tasks from their phone. Hardware manufactured by Dell, HP, Lenovo, Acer, and Asus. (Source: TechCrunch / The Verge)


California Governor Newsom Signs 7 Bills Barring AI Data Centers from Passing Electricity and Water Costs onto Residents. The bills require data centers to pay for grid upgrades and disclose water usage, and the California Public Utilities Commission will establish a new rate class for data centers. (Source: The Verge / Los Angeles Times)


Germany's Far-Right AfD Wins State Parliament Election Again. This is AfD's second state election victory this year, as traditional parties continue to lose voter support. (Source: Axios)


Oura Launches IPO at $2.2 Billion Valuation. The smart ring company's listing is one of the few IPOs in the 2026 consumer hardware space with a pricing signal. (Source: TechCrunch)


Welcome to join the official BlockBeats community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Official Twitter Account: https://twitter.com/BlockBeatsAsia