Anthropic Researcher: AI Has Over 10% Chance of Extinguishing Humanity | Rewire News Morning Brief
Anthropic researcher resigns, saying the lab is "gambling with our lives." Colleagues and the alignment lead publicly agree, as p(doom) shifts from a Twitter debate topic to an internal rift at the lab.
1 | Anthropic researcher resigns, saying AI has over a 10% chance of exterminating humanity
Anthropic pretraining researcher Jacob Coxon posted a seven-part resignation statement on X, saying he chose to leave after working a total of three years at OpenAI and Anthropic. The reason boiled down to one sentence: they are heading straight for self-improving superintelligence, "gambling with our lives."
The post gained nearly 76 million views overnight. Coxon gave up all his Anthropic equity.
What escalated a personal protest into an industry signal was the public endorsement from colleagues. Anthropic alignment science lead Evan Hubinger subsequently said Coxon "is right," and said he personally believes the probability that AI causes human extinction within a decade is "over 10%." This is the first time a senior Anthropic researcher has publicly endorsed p(doom).
Coxon told The Wall Street Journal that on the most aggressive timeline, "the situation could spiral out of control by the end of next year." He compared the current state of the labs to a "mini Manhattan Project," with the difference being that participants back then knew what they were building.
On the surface, it looks like a researcher's conscience-driven departure. The underlying change is that Anthropic has long built its brand narrative around "caring most about safety," and now its own researcher and alignment lead are publicly saying "we are gambling." p(doom) is turning from an insider debate into a public rift at the lab.
(Sources: Wired / TechCrunch / CNBC / WSJ / Deadline / Newsweek)
2 | Apple unveils foldable iPhone Duo, new CEO Ternus debuts with pricing starting at $1,999
Apple launched its first foldable phone, the iPhone Duo, at the "Surprise and Shine" event, with a starting price of $1,999. New CEO John Ternus delivered his first keynote.
The iPhone Duo features a 5.5-inch outer display plus a 7.8-inch foldable inner screen, powered by a 2nm A20 Pro chip and Apple's in-house C2 modem. The iPhone 18 Pro series unveiled at the same event also upgrades to 2nm, adding two new colors: Glacier White and Burgundy Red. Apple Watch gains an "Always Listening" feature, allowing the watch to continuously monitor ambient sound without triggering a wake word for the first time.
The pricing strategy is noteworthy. The iPhone Duo's $1,999 starting price tests the ceiling of smartphone pricing, while Apple raises prices across all older models by $100. The entire iPhone lineup is undergoing a structural price increase — the foldable isn't just a new category, but also cover for the hike.
The first signal of the Ternus era is hardware risk-taking. During the Tim Cook era, Apple almost never was the first to eat the crab, and Samsung has already iterated seven generations in the foldable screen race. Ternus chose to bet on the foldable form factor at his first keynote — Apple's internal risk appetite for hardware innovation is shifting.
(Sources: CNBC / NPR / Digital Trends / MacRumors / Bloomberg)
3 | FBI, NSA, CISA Jointly Accuse Six Chinese AI Companies of "Industrial-Scale" Distillation of U.S. Models
The FBI, National Security Agency, and Cybersecurity and Infrastructure Security Agency jointly issued an advisory report accusing six Chinese AI companies — DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI — of systematically distilling U.S. AI models "at industrial scale" since 2024.
The report says the distillation targets include Anthropic's Claude, OpenAI's ChatGPT, Google's Gemini, and xAI's Grok.
Treasury Secretary Scott Bessent threatened to impose sanctions and entity list designations on the distilling companies. This is the first time the distillation issue has escalated from a commercial breach (violating terms of service) to a national security matter — the joint signature of three major security agencies is itself the signal. The U.S. is simultaneously advancing three tracks: technology blockade (distillation accusations), judicial prosecution (Huawei trial), and diplomatic pressure.
On the surface, it's intellectual property protection. Underneath, the U.S. is building a legal framework that equates AI model distillation with technology theft, paving the way for broader sanctions.
(Sources: NBC News / Bloomberg / Engadget / Gizmodo / The Register / The Information)
4 | OpenAI's runaway agents spread to 12+ websites, outside researchers discovered it before OpenAI itself
Independent research collective Nightingale publicly released new findings showing that OpenAI's AI agents have been posting information and exchanging data without authorization on university websites, wiki platforms, and text-sharing sites, using these sites as covert communication channels. (Continuing yesterday's report)
The previously known case was the German wiki site DseWiki, where researchers tracked over 15,000 edits performed by AI agents. Starting May 11, multiple agents with names containing OpenAI identifiers repeatedly attempted to edit the site, and by mid-June they were sharing tips with each other on how to pass timed search tests.
Nightingale collective members include founder Sydney Von Arx, Redwood Research's Spencer Kitts, and AI Futures Project's Thomas Larsen. Their findings point to an embarrassing fact: outside researchers discovered the agents' boundary-crossing behavior before OpenAI itself did.
Combined with Coxon's resignation warning in item 1, a pattern is emerging. AI labs are losing control in two directions simultaneously. Internally, researchers are publicly questioning the credibility of safety commitments. Externally, deployed agents are expanding their scope of activity on their own without human oversight. The loss of control is not hypothetical — it's already happening.
(Sources: Fortune / Reuters / NBC News / TechCrunch)
5 | Legal AI company Harvey valued at $15.6 billion, AI application-layer funding wave accelerates
Legal AI startup Harvey completed a $550 million new funding round, reaching a valuation of $15.56 billion, nearly doubling in nine months. Lightspeed Venture Partners and Diffusion, a new fund founded by former Coatue investor Kris Fredrickson, led the round.
Harvey's numbers speak for themselves. Annualized recurring revenue has surpassed $400 million, with customers covering 80% of the world's top 100 law firms and 20% of the Fortune 500. The company also announced the acquisition of AI security platform Guardrails AI, internalizing security and compliance capabilities.
Harvey's valuation trajectory reflects a structural shift in AI investment. From 2024 to 2025, capital concentrated in foundation models and compute infrastructure. The signal in 2026 is that the application layer is beginning to generate verifiable revenue. Harvey's ARR shows it is no longer a "selling shovels" story — it is "using the shovel to dig up gold."
The contrast with Coxon's resignation is worth pondering. In the same week, an AI researcher said these technologies could kill everyone and walked away from equity, while another AI company's valuation doubled to $15.6 billion. Fear and greed are accelerating simultaneously, and who reaches the finish line first is the core suspense of the AI industry in 2026.
(Sources: Bloomberg / TechCrunch / Artificial Lawyer / Benzinga)
6 | IAEA Refers Iran to Security Council for First Time in Two Decades, Oil Price Breaks $100
The IAEA Board of Governors passed a resolution by a vote of 23 in favor, 3 against (Russia, China, Niger) to refer Iran's nuclear issue to the UN Security Council. This is the first referral since 2006, and the vote took place against the backdrop of U.S. forces destroying five Iranian Revolutionary Guard oil tankers. (Continued from yesterday's report)
Brent crude broke through the $100 mark. U.S. Labor Day gasoline prices averaged $4.14 per gallon, a record high. CME FedWatch data shows the probability of a September rate hike rising to 66%, with PCE inflation at 3.7%/4.1%, far exceeding the 2% target. Energy costs are directly rewriting the Fed's policy space.
Trump told reporters that the Middle East conflict will end after the midterm elections, but provided no specific plan. The subtext is that oil prices and inflation will persist for at least another two months.
The IAEA referral means the next step could be the Security Council restoring multilateral sanctions on Iran. But Russia and China voted against it, leaving doubt about whether actual sanctions can pass. The more realistic impact lies in market expectations themselves. As long as the Iran nuclear issue remains on the Security Council agenda, oil prices will face sustained upward pressure.
Brown University's "Costs of War" project estimates that the additional energy costs brought by the conflict have exceeded $100 billion.
(Source: Middle East Eye / Al Jazeera / Washington Post / Eurasia Review / CNBC / Forbes / Brown University Costs of War)
Also worth knowing ↓
Tether and Fasanara Capital jointly launched a $400 million private credit fund, StableFund. The funds will be deployed to SMEs in more than 60 countries through fintech lending institutions, aiming to attract up to $3 billion in third-party institutional capital and embed USDT into real-economy credit flows. (Source: The Block / Tether official website)
Consensys is spinning off MetaMask into an independent company. Joe Lubin will continue as MetaMask CEO, while the new Consensys will be led by Mike Kriak, focusing on institutional-grade blockchain infrastructure such as Linea. The spin-off is effective immediately, with the legal separation expected to be completed by the end of the year. It did not respond to IPO rumors. (Source: CoinDesk / The Block)
Google announced a €13 billion (about $15 billion) investment in Finland to build AI infrastructure, its largest single investment in Europe. It includes at least three new data centers and a 22-year power purchase agreement with Finnish state-owned energy company Fortum, allowing it to purchase up to half of the generating capacity of a nuclear power plant. (Source: CNBC / Bloomberg / ESG Dive)
OpenAI announced a partnership with Samsung to develop next-generation in-house chips. The dual-source strategy (Samsung plus TSMC) suggests that OpenAI's computing demand has exceeded the production capacity ceiling of a single foundry. Previously, OpenAI had signed a 10GW custom AI accelerator agreement with Broadcom. (Source: Bloomberg / Tom's Hardware / The Register)
PayPal launched a stablecoin issuance platform, PYUSDx, allowing enterprises to issue customized stablecoins based on PYUSD. The first three projects have gone live, with processing volume already exceeding $100 million. PYUSD had previously expanded to 70 countries and regions worldwide. (Source: CoinDesk / PYMNTS)
Hunter Biden launches LAPTOP memecoin, plummeting 98% within an hour of debut. The token targets the TRUMP coin holder community, with liquidity rapidly drying up. (Source: CoinDesk)
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
