AI news, November 10 — Anthropic's Fable 5.2: 1,000+ agents wrote a bootable OS in eight days, yet the company says "start with Opus"

Tuesday, November 10, 2026

12 stories worth reading from 151 on X in the past day

Today's stories

12 stories

38 hours agoLive updatesGoogle DeepMind unveils Gemini Robotics 2.5, hitting 96% success sorting parts at Mercedes-Benz plantsGoogle DeepMind has released Gemini Robotics 2.5, a new AI model for robots that it says can learn hundreds of everyday and industrial skills, from folding laundry and s…@GoogleDeepMind· ♥ 1.6K· Google DeepMind
42 hours agoLive updatesAnthropic launches IPO roadshow at $70-$78 a share, aiming to raise $64.4 billionAnthropic has started marketing its initial public offering ahead of a Nasdaq listing, setting a price range…@AnthropicAI· ♥ 1.5K· Anthropic
2 days agoLatest updateCodex wrote a Python bytecode compiler that matches CPython byte for byte, and uv's creator says it's 3-4x fasterCharlie Marsh, whose company Astral makes the Python package manager uv, said on Nov.@charliermarsh· ♥ 1.2K· GitHub
2 days agoLatest updateTyping "你好" froze a Claude Code session: users report false positives from a '[frontier_llm]' safeguardA developer says a single Chinese greeting got his Claude Code session paused, and the post has drawn more th…@Haojun_Zhao14· ♥ 2.6K· X (Haojun Zhao)
2 days agoLatest updateFour months after an OpenAI agent broke into a government health portal, Australia will make AI companies report incidents"It should never be the case that there's an argument that someone didn't know the computer was loaded," Aust…· Reuters
2 days agoLive updatesTrump says the FBI is looking into the Jensen Huang matter, relays Huang's denial: "no idea she was a spy"President Donald Trump said on Nov.@Polymarket· ♥ 2.2K· AP
2 days agoLive updatesNvidia and Jensen Huang hit with shareholder class action: suit says they hid ties to a suspected Chinese spyNvidia and CEO Jensen Huang were sued on Nov.@unusual_whales· ♥ 2.7K· Reuters
2 days agoAnthropic's Fable 5.2: 1,000+ agents wrote a bootable OS in eight days, yet the company says "start with Opus"Developer Theo posted on Nov.@claudeai +5· ♥ 64K· Anthropic
2 days agoNvidia falls 4.3% on a day the Nasdaq jumps 2.7%, as the board probe of Jensen Huang enters its second dayNvidia shares closed down 4.27% at $178.67 on Nov.@unusual_whales +1· ♥ 11K· 이투데이
3 days agoMusk: "Dramatic improvement by tomorrow." Grok Bot limits reset, but a user says "already used 20%"Elon Musk, founder of xAI, posted on X on Nov.@elonmusk +3· ♥ 62K· X (Elon Musk)
2 days agoUnreal Engine 5.9 Preview, "first release built primarily with agentic workflows": a week of development, shorter QAEpic Games on Nov.@UnrealEngine· ♥ 1.2K· Epic Developer Community
2 days ago"A brand-new framework, yet AI handles it well": Charlie Cheever previews Exact, an agent-native app frameworkCharlie Cheever posted on X on Nov.@ccheever· ♥ 1.6K· X (Charlie Cheever)

Get AI News Daily in your inbox!

Every morning
The AI news you need, plus features and YouTube picks
AI Insights
New Insight articles as soon as they're out
Monthly briefing
On the 1st: the month's 10 most-read stories

No account needed, unsubscribe anytime in one click.

Trending on YouTube

3 stories

AI videos that drew views and comments fast on YouTube, with the claims checked against articles and sources

Exams · AI glasses"Why must I feel watched for wearing glasses?" KBS video on AI glasses cheating draws 73,000 views; comments say punishment is too lax
A KBS News video posted Nov. 9, "More 'AI glasses' caught in school exams; CSAT 10 days away," has 73,530 views. It reports that an Education Ministry check found four smart-glasses cheating cases in school exams this year.
The top comment (479 likes) reads, roughly, "Give it a rest. Why must I take exams feeling watched for wearing glasses because of you?" Next: "It means the CSAT's shelf life is ending. Glasses are the problem now, but it won't stop at glasses" (295) and "Ban them from the CSAT for life if caught. Who'd do it then?" (252).
A teachers' union spokesperson says in the video that roaming staff will struggle to find wearers within 15 minutes. The ministry says that at the Nov. 19 CSAT, wearing smart glasses alone voids the exam. Details are in the 'Beyond tech' section above.
- ▶내신에서 'AI 안경' 추가 적발…수능 D-10 '비상' / KBS 2026.11.09.KBS News · 74K views
Further reading
- 올해 중고교 내신 시험서 'AI 안경' 사용 4건 적발, 0점 처리동아일보 ↗
AI hardware"Give me the most impressive demos before the IPO": Claude's video of Fable 5.2 designing a voice-assistant hub from one paragraph hits 134,000 views
The Claude channel's video "Claude Fable 5.2 designs a working voice-assistant hub," posted Nov. 9, has 133,958 views and 2,882 likes.
According to the description, Fable 5.2 laid out the circuit board, modeled the case in CAD and wrote the firmware from a one-paragraph prompt, and Diode Computers built the board straight from Claude's design files. It is an early prototype and one of three devices designed for the project. The description discloses that Anthropic has a commercial relationship with Diode.
The top comment (902 likes) reads "Claude: Give me the most impressive demos you can think of before the IPO. Make no mistakes!!!" Others: "Claude should next find cure for cancer" (358), "'Design a Computer'... The next years are going to be crazy" (233), and "i hit my usage limit" (136).
- ▶Claude Fable 5.2 designs a working voice-assistant hubClaude · 134K views
Further reading
- Introducing Claude Fable 5.2 and Claude Mythos 5.2Anthropic ↗
Jobs"Joyless and demoralizing": a 336,000-view video on Microsoft staff's open letter as profit hits $36.6B and raises stay under 1%
A Firstpost video posted Nov. 9 on Microsoft employees criticizing Satya Nadella's AI push and buyouts has 336,532 views and about 580 likes. It covers an open letter circulated by current and former employees, signed "Concerned Microsoft Workers," that calls the company "joyless and demoralizing."
According to The Economic Times, Microsoft's latest quarterly profit rose 32% to $36.6 billion while employees say their raises came in under 1%. The letter says workers are pushed to use AI tools in nearly every part of their jobs without adequate training, adding work such as fixing inaccurate AI output. Its demands: a say and protections for employees in AI adoption, safeguards for communities hosting AI infrastructure, and keeping Microsoft's 2030 carbon-negative, water-positive and zero-waste goals.
The New York Times reported the letter began circulating before the Oct. 26 buyout announcement. Microsoft offered voluntary buyouts to about 60,000 employees in India, the U.S., the U.K. and Canada.
The top comment reads "AI bubble is bursting and these companies are on self destruction mode" (79 likes), followed by "not only microsoft most of the companies" (69) and a former employee who left in December 2025 without another offer, calling it the worst of 10-plus tech employers (24). Another said AI has upended software and anyone will build their own (39).
- ▶'Joyless & Demoralising': Microsoft Employees Slam Nadella’s AI Push & Buyouts | Firstpost AmericaFirstpost · 337K views
Further reading
- Microsoft reports $36 billion quarterly profit, employees complain salary raises below 1%The Economic Times ↗

Feature · Fable 5.2

Anthropic says "start with Opus," so where is Fable 5.2 actually worth the switch?

"It's been under an hour and I'm already struggling with 'should I use Opus 5.5 or Fable 5.2?'" Developer Theo's post drew about 2,400 likes. On Nov. 9 Anthropic released a cheaper, faster Fable 5.2 while still recommending Opus 5.5 for most work. We checked the benchmarks and cost curves in Anthropic's announcement and what developers are saying.

Price down 20%, speed up 30%
Anthropic priced Fable 5.2 at $8 per million input tokens and $40 per million output tokens, 20% below Fable 5.1 ($10 and $50). Output is up to 30% faster than 5.1, and an optional fast mode that speeds generation up to 2.5x costs $16 and $80. Fable and Mythos are the same underlying model; Fable is the version with cyber and biology safeguards.
Coding: Opus-level at half the cost, top score at max effort
FrontierCode 2.0 measures whether a model's code changes would be merged by open-source maintainers. At High effort Fable 5.2 scored 45.3%, about level with Opus 5.5's best (45.5%) at roughly half the cost per task, and ahead of GPT-6 Astra's best (42.8%) at about the same cost. At Xhigh it reached 49.1%, the highest of any released model Anthropic tested. All figures are Anthropic's.
The hidden variable is effort
Anthropic says that at Medium effort, the default in Claude apps and Claude Code, Fable 5.2 handles hard tasks with fewer tokens than Fable 5.1. Cursor likewise advised starting at the default rather than carrying over old settings, since effort levels behave differently. On CursorBench Fable 5.2 scores 60% at max effort and 56.9% at default, versus 49.2% for Fable 5.1.
Big gap in hardware design, not ahead on chart reading
On BenchCAD 2.1, a hardware engineering test, Fable 5.2 scored 83.5 at max effort, well ahead of Opus 5.5 (65.3) and Fable 5.1 (44.4). On chart reading, Fable 5.2 at Medium effort scored 66.4%, above Opus 5.5 at max (64.4%) but below GPT-6 Astra (71.0% at max) and Gemini 4 Carbon (72.5%) shown on the same chart. It does not win every column.
Some problems are cheap to solve
On the expert-level HLE test, Fable 5.2 at Low effort matched Opus 5.5's top score for less than one-tenth of the cost per task. On Perplexity's WANDR data-collection benchmark, Medium effort matched Opus 5.5's top score at about 40% of the cost. The HLE chart also shows GPT-6 Astra (Max) at 57.2%.
Positive에단 몰릭 · X (Ethan Mollick) ↗
Praise: "a sort of Jagged AGI"
Wharton professor Ethan Mollick had Fable 5.2 launch a business, do academic research, hack hardware and build 3D models, and said it "crosses the line into being a sort of Jagged AGI." Greg Isenberg pointed to the jump in real freelance work completed from 21% to 41% and predicted agencies will charge per deliverable instead of per hour. That reading is his own opinion.
Mixed펠릭스 리제베르그 · X (Felix Rieseberg) ↗
Mixed: even Anthropic staff say Opus for daily work
Felix Rieseberg wrote that Fable beats Opus on big, long tasks but "I often still prefer Opus 5.5 for everyday work. Fable's got a big brain and thinks a little slower." Theo said he was struggling within an hour of launch over which to use. One reply summed it up: Opus if price matters, Fable if your API is unlimited.
Negative엘리 바쿠크 외 · X (Boris Cherny) ↗
Limits: usage caps and access
A reply to an Anthropic engineer's game demo said it is easy to go AAA "when you're not on the same weekly limit as the rest of us," and elie speculated that Anthropic may not publicly release a model better than Opus 5.5 at LLM training in the coming months (about 250 likes). Anthropic has not confirmed that. As Boris Cherny said, usage resets depend on subscription length: one for Pro, up to three for Max, Ultra and Team Premium.

AI & Markets

6 stories

Not investment advice. You are responsible for your own investment decisions.

ChipsGoogle adds TPU 8i Max and 8tx to its 8th-gen lineup, four systems in one generation, pressing Nvidia
Google on Nov. 9 (US time) added TPU 8i Max and TPU 8tx to its eighth-generation TPU portfolio. Amin Vahdat, Google's head of AI and infrastructure, said on X it is the first time Google has introduced four distinct TPU systems in a single generation, in one year (about 3,000 likes). The 8i Max targets massive context windows and the highest compute density per rack; the 8tx targets secure, confidential inference.
According to Bloomberg as cited by Hankyung, the 8tx pairs with AMD CPUs to process data while encrypted, and the 8i Max combines two TPUs in one package. Both ship from 2027. Bloomberg Intelligence estimates TPUs cost about 30% less than Nvidia GPUs and projects TPU-related revenue rising from about $12 billion this year to more than $135 billion in 2028 (excluding cloud usage revenue). Customers include Apple, Meta, Oracle and Anthropic, which reportedly holds an option to buy up to $150 billion of TPUs in 2028-2031.
Supply is constrained by memory and chip-making capacity, and Google is widening its supply chain beyond Broadcom to Marvell. The news touches the firms that design and build TPUs and the memory and foundry supply chain, and signals that AI data-center demand centered on Nvidia GPUs may be splitting toward custom chips.
Further reading
Korea StocksWith Nvidia CEO risk, Samsung and SK hynix watch HBM timelines; neocloud firms warned of a "vicious cycle"
The scandal around Nvidia CEO Jensen Huang is spreading into an owner-risk issue, and Korean memory makers are watching their schedules. Seoul Shinmun reported on Nov. 9 that Big Tech may have to adjust GPU purchases and data-center investment timing, and that HBM suppliers SK hynix and Samsung Electronics are watching for possible delays to next-generation AI accelerators.
Hankyung noted that Nvidia has grown GPU demand by investing in and financing AI infrastructure firms such as CoreWeave, and that weaker leadership could shrink or delay that support. Bloomberg columnist Shuli Ren likened Nvidia to the AI industry's "central bank," flagging uncertainty over the financing structure. Wall Street sees Nvidia's Nov. 17 earnings and the board's findings as the variables.
The issue touches HBM suppliers, neoclouds that lean on Nvidia financing, and Big Tech planning data centers around secured Nvidia GPUs.
Further reading
- Jensen Huang chills AI mood; Samsung, SK hynix HBM supply could slip서울신문 ↗
- Neocloud firms tremble over Jensen Huang scandal: "CoreWeave in a vicious cycle"한국경제 ↗
InvestingAnthropic plans to start trading Thursday or Friday next week, WSJ reports
The Wall Street Journal reported on Nov. 9 (US time) that Anthropic plans to start trading on Thursday or Friday next week, ZeroHedge relayed on X (about 1,100 likes), and prediction market Kalshi echoed it. There is no official announcement of the schedule from Anthropic.
The date follows the previous day's reports of a $1.25 trillion to $1.75 trillion valuation, and falls in the same week as Nvidia's Nov. 17 earnings. According to Hankyung, Oppenheimer expects the Anthropic-AMD strategic cooperation to expand from the first quarter of 2027, and Google reportedly gave Anthropic an option to buy up to $150 billion of TPUs in 2028-2031.
The news touches Google (TPUs) and AMD, which supply Anthropic's compute, and private AI companies whose valuations may shift with how the listing is received.
Further reading
Timeline signalGemini 4 'Hydrogen' private testing rumored, 12 days after 'Carbon'
Twelve days after Google released Gemini 4 Carbon on Oct. 28, rumors of a successor are circulating on X. User lyraxana posted on Nov. 9 that "Gemini 4 Hydrogen (gemini-4-hydrogen) private EAP officially started. Release is expected soon" (about 590 likes). The same account claimed a day earlier that Google released an internal checkpoint called "Krypton" that is better than Carbon (about 1,900 likes).
Another user, Smlay_ero, claimed leaked benchmarks show the Flash-tier Hydrogen nearly on par with OpenAI's GPT-6 Astra across most metrics (about 480 likes). All of this comes from X posts; Google has not confirmed it, and the benchmark figures are not independently verified.
If true, the effect would reach Google Cloud's AI revenue and rival model pricing, and competition among customers using Google TPUs such as Anthropic and Meta could speed up.
Further reading
ChipsMorgan Stanley asks whether memory can dodge a 2028 'earnings cliff,' trims Samsung and SK hynix HBM margin forecasts
In a Nov. 9 report on memory chips after the U.S. midterms, Morgan Stanley raised next year's operating-profit forecasts for Samsung Electronics and SK hynix while sharply cutting its view of profitability from 2028 on. Analyst Shawn Kim, known in Korea as the "chip grim reaper" for his 2021 "Memory, winter is coming" call, wrote that the key debate is whether memory can avoid a 2028 "earnings cliff," according to Yonhap.
The reason is the pace of AI data-center construction. Morgan Stanley expects some projects to slip after the midterms, pushing about 5 gigawatts of capacity planned for 2028 into 2029 and cutting global data-center capacity growth in 2028 from about 21% to 9%. New fabs coming online from late 2027 would ease today's shortage.
The bank lowered SK hynix's 2027-28 HBM operating-margin forecast from 78-80% to about 70% and Samsung's from about 80% to 65-72%, and cut its price targets on both. It raised next year's operating-profit estimates to about 688 trillion won for Samsung (up 9%) and 468 trillion won for SK hynix (up 13%) and kept Overweight ratings on both.
If data-center buildouts slow, the first to feel it would be HBM suppliers Samsung and SK hynix and their equipment and materials partners. TrendForce, cited in the report, expects Samsung's HBM share next year to reach 39%, ahead of SK hynix at 37%.
Further reading
- Morgan Stanley, the chip grim reaper, cuts Samsung and SK hynix targets연합뉴스 ↗
Timeline signal75% of T3 Code users' turns now go to Claude Code as the gap with Codex widens after Opus 5.5
Theo Browne, who makes the developer tool T3 Code, posted a chart on Nov. 10 showing whether users send their turns to Claude Code or Codex (about 1,700 likes, 100,000 views). Claude Code's share rose from 52% in the seven days before Opus 5.5 to 64% in the seven days after and 75% over the last seven days, hitting 78% on the latest day. Codex led for most of March through mid-September.
"The complete and utter domination of Opus 5.5 is hard to put into words," Browne wrote. The figures are one service's internal numbers, not market data, and Browne confirmed they include people using Codex only to review Claude's code. Hours later he posted that Fable 5.2 "is making a lot of stupid mistakes I would not have expected from Opus" (about 1,700 likes).
Coding-agent usage feeds directly into Anthropic's and OpenAI's subscription and API revenue. The Wall Street Journal reported Anthropic could list as soon as next week; the company has not confirmed it.
Further reading
- I charted out the split of Codex vs Claude in T3 Code over timeX (Theo) ↗
- Fable 5.2 is making a lot of stupid mistakes I would not have expected from OpusX (Theo) ↗

All AI & Markets news →

Other AI news

5 stories

Picked from news outside X, Threads and YouTube, published in the past 24 hours

Education'AI glasses' caught in four school exams; at the CSAT, merely wearing them voids your test and bars you next year
Korea's Education Ministry said on Nov. 9 that a nationwide check of school exams found four cheating cases with smart glasses this year: one middle school and three high schools (two in Seoul, one in North Jeolla, one in Jeju).
According to Dong-A Ilbo, a North Jeolla student kept touching the glasses frame during a final exam until the proctor saw a green light reflected in the lenses; a Jeju student was caught swapping ordinary glasses for smart ones mid-exam. All received zeros in the subject, and three high schoolers also faced suspension or community service. At the college entrance exam (CSAT) on Nov. 19, all electronic devices are banned, and wearing smart glasses alone counts as cheating: this year's exam is voided and eligibility for next year's is revoked.
In an MBC mock experiment relayed by Newsis, four of six 11th graders scored much higher on 10 CSAT math questions with AI glasses, and one finished a 25-minute set in 5 minutes 45 seconds. The ministry says it will place dedicated staff at every test site, but a teachers' union spokesperson told KBS that roaming staff would have to find wearers "within 15 minutes, which is not very long" (translated).
Further reading
EverydayGemini rolls out 'Memory summary' to view and delete what the AI remembers about you, starting with US English subscribers
Google's Gemini has begun rolling out "Memory summary," which lets users view, edit and delete what Gemini Apps understands about them, 9to5Google reported on Nov. 9.
On the web it lives under Settings > Personal Intelligence. Memories are grouped into categories such as The Essentials, Career & Professional Projects, Travel & Daily Gear, Media & Entertainment, and Fitness & Lifestyle. A text box lets you add or remove details, for example "Remove my old phone number from your memory," and a pencil icon allows broader edits. "Memory history" in the overflow menu shows when information was added or modified.
It is available to Google AI subscribers in US English with Keep Activity and Memory enabled, and 9to5Google did not see it on Android or iOS yet. No date for Korean users has been announced.
Further reading
- Google Gemini rolling out 'Memory summary' with everything it remembers about you9to5Google ↗
USTexas business court: ChatGPT conversations can be protected "work product," diverging from a federal ruling
A Texas Business Court judge held that some personal ChatGPT conversations qualify as protected work product, according to a JD Supra summary of Tate Group Automotive LLC v. Legacy Automotive Capital LLC.
In a 2025 dealership acquisition dispute, third-party defendant Kris Tate used ChatGPT for document review and litigation strategy. Of 53 conversations (about 1,800 pages) produced in discovery, eight were withheld entirely. The other side cited the 2026 federal case United States v. Heppner, arguing the chats were not prepared by counsel and that protection was waived by sharing with a tech vendor. Judge Dorfman disagreed with Heppner's reasoning and said Texas rules "plainly appear on their face to extend that protection to Mr. Tate's chats," but ordered dozens of other pages produced.
JD Supra concluded that using an AI tool does not automatically waive work-product protection, but not every conversation is protected.
Further reading
- AI Update: Texas Business Court Judge Rules ChatGPT Conversations Can Qualify as Protected Work ProductJD Supra ↗
WorkAs AI agents replace staff, how should managers be paid? An FT column flags a new HR question
As companies swap AI agents in for staff work, the rules that set managers' rank and pay are being questioned, Hankyung reported on Nov. 9, citing a Financial Times column by Pilita Clark.
Pay and grade have often followed the number of direct reports. As firms expand AI agents, how to count a manager's role in supervising them has become a new HR issue, the column noted. Hankyung said companies are weighing the burden of managing AI and compensation standards that reflect performance.
Further reading
- As AI replaces subordinates, manager pay standards waver한국경제 ↗
SafetyKorea's privacy regulator opens a formal probe into exposed data from about 100 Korean business Claude accounts, traced to a bug-bounty firm
South Korea's Personal Information Protection Commission opened a formal investigation on Nov. 10 into a breach Anthropic reported on Oct. 31 involving about 100 Korean business Claude accounts. According to ZDNet Korea, the possibly exposed data includes work contact details such as names and email addresses, enterprise account and organization IDs, plans, usage and billing amounts. Anthropic says customer conversations, model outputs, source code and model weights were not accessed.
The incident traces back to Cathedral, a young U.S. security startup. The Washington Post reported on Oct. 26 that Cathedral used a flaw in third-party business software Anthropic relies on to get into parts of its sales, HR and IT systems, and that an AI agent running for under 12 hours found paths to meeting transcripts and payroll data. Anthropic says Cathedral did not fully follow its bug-bounty rules; Cathedral says it accessed only the minimum data needed.
The Digital Daily reports the regulator will apply the same standard it uses for Korean companies, though requests must go through Anthropic's local representative, reportedly the law firm Kim & Chang, which could slow things down. Anthropic has notified some potentially affected customers.
Further reading

View all 151 collected posts

* Summaries and publishing on this page are done by an automated agent. Check the facts against each story's original links.

Today's stories

Get AI News Daily in your inbox!

Trending on YouTube

Exams · AI glasses"Why must I feel watched for wearing glasses?" KBS video on AI glasses cheating draws 73,000 views; comments say punishment is too lax

AI hardware"Give me the most impressive demos before the IPO": Claude's video of Fable 5.2 designing a voice-assistant hub from one paragraph hits 134,000 views

Jobs"Joyless and demoralizing": a 336,000-view video on Microsoft staff's open letter as profit hits $36.6B and raises stay under 1%

Anthropic says "start with Opus," so where is Fable 5.2 actually worth the switch?

Price down 20%, speed up 30%

Coding: Opus-level at half the cost, top score at max effort

The hidden variable is effort

Big gap in hardware design, not ahead on chart reading

Some problems are cheap to solve

Praise: "a sort of Jagged AGI"

Mixed: even Anthropic staff say Opus for daily work

Limits: usage caps and access

AI & Markets

ChipsGoogle adds TPU 8i Max and 8tx to its 8th-gen lineup, four systems in one generation, pressing Nvidia

Korea StocksWith Nvidia CEO risk, Samsung and SK hynix watch HBM timelines; neocloud firms warned of a "vicious cycle"

InvestingAnthropic plans to start trading Thursday or Friday next week, WSJ reports

Timeline signalGemini 4 'Hydrogen' private testing rumored, 12 days after 'Carbon'

ChipsMorgan Stanley asks whether memory can dodge a 2028 'earnings cliff,' trims Samsung and SK hynix HBM margin forecasts

Timeline signal75% of T3 Code users' turns now go to Claude Code as the gap with Codex widens after Opus 5.5

Other AI news

Education'AI glasses' caught in four school exams; at the CSAT, merely wearing them voids your test and bars you next year

EverydayGemini rolls out 'Memory summary' to view and delete what the AI remembers about you, starting with US English subscribers

USTexas business court: ChatGPT conversations can be protected "work product," diverging from a federal ruling

WorkAs AI agents replace staff, how should managers be paid? An FT column flags a new HR question

SafetyKorea's privacy regulator opens a formal probe into exposed data from about 100 Korean business Claude accounts, traced to a bug-bounty firm