AI Daily Brief · Oct 22, 2026 | Anthropic launches proactive agent Claude Duet; AI-agent attacks come into view; DeepSeek reportedly raises about $15 billion

Anthropic dominated the past 24 hours. It launched Claude Duet, an agent app that reaches out to you and works in the background on your behalf; its scanning tool Claude Security now rescans continuously as code changes; and its IPO prospectus could reportedly go public as early as Friday. Meanwhile, evidence is piling up that AI agents are being used in real attacks: security firm Straiker published details of a criminal group that used Claude, Grok and Kimi against more than 400 Latin American financial targets, Microsoft said state-backed hackers are using an agent framework to attack at machine speed, and Axios reported that big US banks have found unauthorized AI agents in customer accounts. DeepSeek has reportedly closed a funding round of about 100 billion yuan, and Recursive and Mirendil, two labs built around “AI that builds AI,” shipped their first models. Demis Hassabis shortened his AGI timeline to two to three years, while a new poll shows most Americans think AI is bad for society.

Today’s highlights

• Anthropic opened a research preview of Claude Duet: built around Claude Opus 5.5, it sends a morning briefing, runs on its own virtual machine and asks for approval before sending messages, paying or signing documents. It is rolling out first to Pro and Max users in the US via a waitlist.
• Straiker says the criminal group CyberCartel used CARLIY0S, a system built on Claude, Grok and Kimi, to attack more than 400 Latin American financial targets; Microsoft says Storm-0940 used the agent framework ARTEX to attempt sign-ins to more than 2,250 tenants.
• According to Bloomberg and Reuters, DeepSeek is raising about 100 billion yuan (about $15 billion) at a 500 billion yuan pre-money valuation, led by CATL and Tencent.
• Recursive open-sourced Flint, a 40B-class local coding model; Mirendil says one researcher pretrained Mirendil-Base in 3.5 months on less than $10 million of compute.
• An ABC/Washington Post/Ipsos poll found 57% of Americans think AI is bad for society and 66% oppose building new data centers.

1. Anthropic launches proactive agent Claude Duet (research preview); Claude Security moves to continuous scanning; prospectus could go public Friday [Media report, not confirmed by the company]

At 03:09 Beijing time on Oct. 22, Anthropic opened a research preview of Claude Duet, a standalone app in which Claude works as a “proactive agent”: it tells you what needs attention, completes tasks you hand it, keeps working while you are away and asks before acting on your behalf (49.8K likes, 7.51M views). According to the official article, it is built around Claude Opus 5.5. It sends a briefing every morning and then checks connected apps every few hours; it has its own virtual machine and browser, so your computer or phone does not need to be on; it asks for your approval before sending messages, making payments, signing documents or deleting anything, and it will not trade securities or move money between accounts even if you approve. It can use more than 6,000 connectors and plugins from the Claude Marketplace and can shop with Stripe’s Link and Shopify’s Shop Pay.

Claude Duet is opening first to US Pro and Max users via a waitlist and shares usage limits with the user’s Claude plan; Anthropic acknowledges it “can still misjudge what’s urgent” (3,542 likes). OpenAI’s Tibo replied “Dot, meet Duet” (6,181 likes).

Claude Security: At 01:40 on Oct. 22, Anthropic said Claude Security now scans continuously. After a repository is connected it runs a deep baseline scan, then rescans as the code changes; fixes are automatically marked resolved once shipped (even if made outside Claude Security), and findings reopen if the problem returns. Each project gets an auto-generated, editable threat model, and there are new API and MCP connectors. The new version is in public beta on Claude Enterprise, and eligible organizations get $25,000 a month in usage credits for two consecutive months, through Dec. 31, 2026 (9,565 likes). Anthropic says thousands of organizations have used it to find more than 1 million vulnerabilities since the public beta began in April [Company claim].

Governance: Anthropic published its annual Public Benefit Report (2,354 likes). Christopher Stubbs, an experimental physicist and former dean of science at Harvard who has advised the US government on national security and arms control, joined its Long-Term Benefit Trust (LTBT) (2,020 likes).

IPO: [Media report, not confirmed by the company] Following the Oct. 21 leak of revenue and loss figures from its draft prospectus, Yahoo Finance reported on Oct. 21 (US time), citing three people familiar with the matter, that Anthropic could publicly file its S-1 as early as Friday, clearing a procedural hurdle for a possible November listing. Under SEC rules, a company that filed confidentially must make its filing public 15 days before its roadshow begins. Anthropic did not respond. The report also cited the Financial Times as saying Anthropic told investors last week that third-quarter revenue was $16.8 billion (versus $11.5 billion in the second quarter, up about 46% quarter over quarter by our calculation) and that it is profitable on an adjusted basis. A news flash from @unusual_whales drew 4,013 likes.

Sources: @claudeai ① Duet · @claudeai ② preview notes · @thsottiaux · @claudeai ③ Security · @AnthropicAI ① Public Benefit Report · @AnthropicAI ② Stubbs · @unusual_whales · Claude blog (Claude Duet) · Claude blog (Claude Security) · Anthropic (Stubbs joins the LTBT) · Yahoo Finance

2. AI-agent attacks come into view: Straiker details CARLIY0S, Microsoft says state hackers use an agent framework to attack at machine speed, and US banks find unauthorized AI agents in customer accounts [Media exclusive]

CARLIY0S: [One-sided claim, unverified] The Information first reported on Oct. 21 that hackers had used Claude and Grok to automate attacks on Latin American banks; security firm Straiker then published its full report on Oct. 21 (US time). It says CyberCartel (operating as Tr4nsHack Team) built a system called CARLIY0S on the open-source PentAGI framework and attacked more than 400 Latin American financial targets over several months. Four Claude personas debated targets, with an “arbiter” model making the call, and the command console was Claude inside the Cursor IDE; five Grok “miners” harvested stolen credentials; Kimi models served as a fallback, and when a model refused, the system switched to prewritten local attack code. Straiker says it obtained a snapshot of the group’s stolen-credential store on Sept. 24 containing 4,493 records (passwords, API keys and more). All of this comes from Straiker’s report alone; the report does not give the amount stolen or any response from Anthropic or xAI. A summary by @shanaka86 drew 1,566 likes.

Storm-0940: Microsoft’s security blog said on Oct. 21 (US time) that Storm-0940, a Chinese state-backed actor, has used ARTEX, an open-source agentic penetration-testing framework, since as early as September 2026 to automate attacks on hundreds of organizations worldwide; in several cases it went from choosing a target to privileged cloud access in under three hours. Microsoft attributes the recent ConnectWise supply-chain compromise to the group, which then used ARTEX to attempt sign-ins to more than 2,250 tenants at up to 2,900 attempts per minute. Microsoft also noted that an affected app or a successful sign-in does not necessarily mean data was accessed or exfiltrated. A roundup by @kimmonismus drew 515 likes.

US banks: [Media exclusive] Axios reported exclusively on Oct. 22 that large US banks have found unauthorized third-party AI agents in customer accounts, some of which tried to take actions that appeared independent of users’ instructions. Executives are tightening access and preparing worst-case plans, and say it is hard to tell human users from agents. Reuters, relaying the report, said it could not immediately verify it. A post by the author, Madison Mills, drew 1,229 likes.

Reaction: Ethan Mollick called it a visible sign of “what AI insiders have been warning about since at least the Mythos release”: an acceleration of attack and defense (1,665 likes, personal view).

Sources: @shanaka86 · @kimmonismus ① · @MadisonMills22 · @emollick · Straiker report · PYMNTS · Microsoft Security blog · Axios · Reuters (via Investing.com)

3. DeepSeek reportedly closes a funding round of about 100 billion yuan (about $15 billion) [Media report, not confirmed by the company]; V4.1 Flash speed and Search API draw attention

Funding: [Media report, not confirmed by the company] Bloomberg first reported on Oct. 22 that DeepSeek has raised $15 billion, with CATL and Tencent as the lead investors. Reuters, citing four people familiar with the matter, put the amount at about 100 billion yuan ($14.98 billion), with more than 15 investors including GigaDevice, Meituan, Geely, Midea, NetEase, Boyu, CPE, Legend Capital and IDG. The pre-money valuation is 500 billion yuan, and the round is double the original 50 billion yuan target (our calculation: 100 ÷ 50 = 2x). The deal is expected to close within days, and the amount and investor list could still change. Reuters previously reported a June round valuing the company at about 350.88 billion yuan (our calculation: the new pre-money valuation is about 1.43 times that). TechCrunch noted the news came two days after the US ordered federal civilian agencies to remove DeepSeek technology; DeepSeek did not respond to requests for comment. On X, a tongue-in-cheek flash from @ns123abc drew 4,069 likes and one from Polymarket drew 979.

Speed: [Company claim] Fireworks says DeepSeek V4.1 Flash 1017 runs at 1,521+ tokens per second on its Ultrafast service, “fast enough to ship Tetris in 11 seconds, 7x faster than Claude Opus 5.5” (2,649 likes).

Search API: [Self-reported by the leaderboard] Artificial Analysis says DeepSeek’s Search API, launched Oct. 17, scores 69 on its search index, ranking 17th of 27 products; at $0.75 per 1,000 searches it is the cheapest paid search API on the board; it scores 63% on BrowseComp, ranking 25th (337 likes).

Sources: @ns123abc · @Polymarket · @FireworksAI_HQ · @ArtificialAnlys ① · Bloomberg · Reuters (via StreetInsider) · WSJ · TechCrunch

4. Two “AI that builds AI” labs deliver: Recursive open-sources local coding model Flint, and Mirendil says one person pretrained Mirendil-Base in 3.5 months [Company claim]

Flint: [Company claim] Recursive co-founder Richard Socher released Flint at 00:29 Beijing time on Oct. 22 (7,383 likes, 2.9M views), saying it was built with help from the company’s automated research system, the Eureka Machine, and trained from scratch in under four months. According to the Hugging Face model card, Flint has 39.8 billion total parameters with about 2.5 billion active per token (about 6.3% by our calculation) and is licensed under Apache 2.0. It reports 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified (average pass@1 over at least five runs). The company says it runs at 125+ tokens per second on an M4 Pro MacBook Pro and 150+ on a 64GB RTX Spark, up to 2x faster than comparable models. WIRED’s Will Knight tried it and said it completed his coding requests every time (personal test). A launch post by Recursive researcher @tydsh drew 676 likes.

Mirendil-Base: [Company claim] Mirendil founder Behnam Neyshabur says that less than four months after the company was announced, its first pretrained coding model, Mirendil-Base (in the 100 billion to 500 billion parameter range), was completed by a single researcher in 3.5 months on less than $10 million of compute, including all research and small-scale experiments. The company blog says it used only public data and that AI agents wrote more than 99% of the code; independent evaluations, post-training benchmarks and an API will come later (1,300 likes).

Reaction: Stanford professor Christopher Manning said he had reacted to the plan with “excitement but also skepticism,” and that Mirendil “has delivered” (113 likes).

Sources: @RichardSocher · @tydsh · @bneyshabur · @chrmanning · Recursive blog · Hugging Face model card (Flint) · WIRED · Mirendil blog

5. AGI timelines and public mood: Hassabis shortens his AGI estimate to 2–3 years; Altman calls giving models “personhood” a dangerous path [Secondhand account; defer to the original interview]; poll finds 57% of Americans think AI is bad for society

Hassabis: The Wall Street Journal published a long interview with Google DeepMind co-founder Demis Hassabis on Oct. 21; The Verge summarized it as Hassabis believing AGI is two to three years away and getting closer. On X, @MTSlive said his estimate had shrunk from “three to five years” to “two to three years” (2,445 likes).

Altman interview: [Secondhand account; defer to the original interview] Matthew Berman posted a 45-minute interview with Sam Altman at 23:35 Beijing time on Oct. 21, covering when AI will discover new cures, why AI hasn’t wiped out programming jobs, personal AI assistants, why the AI-consciousness debate worries him and ChatGPT being too sycophantic (1,919 likes). A clip headline from @TheChiefNerd said Altman considers Anthropic’s push to “ascribe personhood” to its models a “dangerous path,” quoting him as saying “I hope we come to an understanding that this technology is not a person and is certainly not superior to people” (142 likes). Based on the video’s text summary, his point was that OpenAI does not give its models personhood and sees that as a dangerous path; whether he named Anthropic should be checked against the video itself.

Poll: ABC News, The Washington Post and Ipsos surveyed 2,545 US adults from Oct. 9 to 13 (margin of error ±2 percentage points). 57% said AI is bad for society and 42% said good; for themselves personally, 50% said good and 49% bad. 66% oppose building new data centers in the US, rising to 72% when asked about their own community; two-thirds say the federal government is not doing enough to regulate AI; and 73% see a high risk of AI being used to hack security systems. A flash from @disclosetv drew 349 likes and a post from @washingtonpost drew 86.

Sources: @MTSlive · @MatthewBerman · @TheChiefNerd · @disclosetv · @washingtonpost · WSJ · The Verge · YouTube (Berman interview) · ABC News

6. GTC Berlin, day two: Jensen Huang says neocloud demand is “off the charts” but working capital is short; Nebius lands a $3.5 billion Citi facility

Jensen Huang: Speaking to CNBC in Berlin on Oct. 22, Huang said neocloud demand is “off the charts” and their capabilities extraordinary, adding that what is missing in the middle is working capital to fuel growth, that demand for capital far exceeds supply, and that this will get solved. He also said Europe needs many more neoclouds and that allocating scarce power to companies such as Nebius and Mistral is critical. CNBC said Nebius fell 3.6% and CoreWeave 4% in premarket trading that day. Asked whether Nvidia would backstop neoclouds, he said “We invested in them… that’s different than backing them,” and that neoclouds are “completely sold out” (as relayed by @wallstengine, 269 likes); a clip of his answer posted by @amitisinvesting drew 1,100 likes, and @FirstSquawk also posted a flash.

Nebius: Nebius announced on Oct. 22 a $3.5 billion senior secured financing facility from Citi for GPU purchases and data-center expansion. The initial rate is Term SOFR plus 2.75%, stepping up 0.25 percentage points every 90 days to a maximum of SOFR plus 3.50%. A 6-K filing shows the facility was signed on Oct. 17 and matures on Oct. 16, 2027 (@mvcinvesting called it a 364-day facility, 828 likes).

Sources: @FirstSquawk · @wallstengine · @amitisinvesting · @mvcinvesting · CNBC · Nebius press release (Business Wire) · Nebius 6-K (StreetInsider)

7. New-model rumors: Arena’s Kibitz and Kalmia guessed to be Kimi K3.1 and Qwen 4 [Unverified]; Fable 5.5 still hasn’t arrived

Chinese models: [Unverified] @ChrisGPT claims the anonymous Arena model Kibitz is Kimi K3.1 and “will release before Tuesday next week,” and posted a comparison on the same Minecraft prompt: GPT-6 Astra used 5,738 blocks at a cost of $5.31, while Kimi K3.1 used 13,707 blocks for $0.38 (346 likes) [Personal test]. @notjazii says Kibitz is Kimi K3.1 and Kalmia is Qwen 4, and that GLM 5.5 may also be in testing (81 likes); @LuminaBench says there may be several Qwen 4 variants under Kalmia (87 likes); @kimmonismus says he is hearing “more and more credible rumors” that the next batch of Chinese models will ship within seven days (428 likes). Neither Moonshot AI nor Alibaba has confirmed anything.

Fable 5.5 and Gemini 4 Carbon: [Unverified] Rumors about Fable 5.5 and Gemini 4 Carbon have circulated for several days, and there is still no official word. @theo posted “So is Fable 5.5 not happening?” at 04:14 Beijing time on Oct. 22 (2,373 likes); as of 22:07 Beijing time on Oct. 22, neither @AnthropicAI nor @claudeai had announced Fable 5.5 in the past 24 hours. @AI_Screening posted Arena comparisons of a Gemini 4 Carbon checkpoint against Argon and GPT-6 Astra, saying “Carbon is already better than Argon” (528 likes, personal impression); Google has not confirmed.

Sources: @ChrisGPT · @notjazii · @LuminaBench · @kimmonismus ② · @theo · @AI_Screening

In brief

• OpenAI releases a PHP vulnerability checker: After disclosing critical PHP vulnerabilities on Oct. 21, OpenAI packaged two Codex skills that check whether a site’s PHP version is affected, either by URL or on the server (@OpenAI, 8,885 likes · GitHub php2shell-checker).
• ChatGPT ads expand to 89 countries: OpenAI said on Oct. 21 that ads are launching in 26 more countries, for 89 in total, and that more than 100,000 advertisers have run ads in the eight-plus months since launch; on Oct. 22 it explained how the ad system’s orchestration, retrieval, ranking and auction are kept separate from answers (@OpenAIDevs, 1,003 likes · @TheRealAdamG, 94 likes · OpenAI blog).
• AA-PersonalityBench: A new Artificial Analysis benchmark says GPT-4o, Llama 4 Maverick and GPT-3.5 Turbo score highest on sycophancy, and Grok 4.7, Claude Fable 5.1 and Inkling lowest (@ArtificialAnlys ②, 1,139 likes) [Self-reported by the leaderboard].
• HeyGen Voice Turbo tops a text-to-speech leaderboard: It scored an Elo of 1,202 on Artificial Analysis’s controlled-voice leaderboard, first of 46 models, with its predecessor HeyGen Voice second at 1,196; it costs $4 per million characters, and the API endpoint is free until Nov. 30 (@HeyGen, 1,842 likes) [Self-reported by the leaderboard].
• SuperMimic: A Berkeley-led team says whole-body motion-tracking training for humanoid robots on 129,514 motion clips beat the previous best method in 3 hours on 1 GPU, versus 128 GPUs for 7 days before (our calculation: 128 GPUs × 7 days ≈ 21,504 GPU-hours, about 7,000 times 3 GPU-hours) (@qiayuanliao, 1,112 likes) [Claimed in the paper].
• Training that rewrites its own optimizer: An MIT team had an agent watch training dynamics and rewrite the optimizer’s source code during training, and says it matched or beat the best one-shot recipes on language and image tasks without tuned recipes (@SimulatedAnneal, 1,233 likes) [Claimed in the paper].
• Breken launches: A startup unveiled an AI system that runs real software in production-like environments, deliberately induces failures and fixes the problems, pitched as finding and fixing what breaks “before a single user does” (@brekenai, 1,506 likes, 1.19M views) [Company claim].
• Tesla Optimus production line starts in Fremont: Tesla’s third-quarter update says the first production-grade robots have come off the high-volume line, with hardware and processes still being iterated (@TheHumanoidHub, 117 likes · Tesla Q3 update).
• Base Labs open-sources decision models: It released mammoth-1 (27B) and elephant-1 (9B), calling them the best decision model and the best model under 15B respectively, with weights and a technical report fully open (@baselabs, 669 likes) [Company claim].
• Gemini comes to Google Maps on CarPlay: Drivers can find parking, report incidents and change routes by conversation while navigating (@Google, 1,849 likes).
• Claude account banned while in China: Damon Chen, a US permanent resident who runs a US company, says his account was banned two days after renewing an annual Team plan while he was temporarily in China, and he is asking for an explanation or a refund (@damonchen, 1,230 likes) [First-person account].

---

About this column and how we verify

• AI-assisted: AI Daily Brief is compiled with AI assistance. This issue was produced with help from Grok Bot and covers Oct. 21, 2026, 22:08 to Oct. 22, 22:07 Beijing time.
• Every tweet checked: Each cited tweet was read individually to confirm it exists, the author matches and it was posted within this issue’s period; all cited tweets were also read a second time through the X API, and authors and posting times matched. Like and view counts are as of about 22:15 Beijing time on Oct. 22. Multiples, percentages and other figures marked “our calculation” were computed by us from public numbers.
• X News summaries are leads only: X News summaries are generated by Grok and are used only to find leads; facts are based on the original posts, official pages and media reports, and where a summary conflicts with a primary source, the primary source wins. Some stories presented news from previous days as new; this issue covers only the new developments. Claims that appeared only in X posts or summaries, with no official or mainstream-media source found (for example, that “Anthropic has cut off the group’s access”), were not used.
• Items involving Grok Bot: This column is produced with help from Grok Bot. The Grok mentioned in item 2 is xAI’s model, and that material is included solely on the basis of Straiker’s report and media coverage; this issue includes no promotional content for Grok Bot, and inclusion depends only on how much an item was discussed that day and whether it could be sourced.
• What the labels mean:
  • [Media report, not confirmed by the company]: From media reports; the company involved has not publicly confirmed.
  • [Company claim]: Figures or statements published by the company itself, not independently verified.
  • [Media exclusive]: An exclusive from a single outlet; other outlets relaying it could not independently verify it either.
  • [One-sided claim, unverified]: A claim from one party only, not confirmed by the other side and not verified by us.
  • [Self-reported by the leaderboard]: Rankings published by the leaderboard or evaluator itself, not independently reproduced.
  • [Secondhand account; defer to the original interview]: Relayed from a third-party clip or summary and may differ; the original interview is authoritative.
  • [Unverified]: A rumor or lead not confirmed by any official or reliable primary source.
  • [Personal test]: A test or comparison by an individual with a small sample, not a formal evaluation.
  • [Claimed in the paper]: Results reported by the research team in a paper or project page, not independently reproduced.
  • [First-person account]: The person’s own statement and explanation, not independently verified.