AI Daily Brief · Oct 28, 2026 | Google DeepMind unveils Genie 4; Anthropic buys a one-person biotech and publishes exploit-chain and expert-judgment studies; AA Intelligence Index v5; Google TPUs come to Oracle Cloud

Over the past 24 hours, world models, AI security research and benchmarking all moved at once. Google DeepMind released Genie 4, adding native audio and real-time natural-language control, and for the first time plans to offer Genie through an API. Anthropic, in a single day, shared drug-development progress and its acquisition of one-person biotech Mithril, and published two studies: an unreleased internal model built five exploit chains across Chrome, Safari, Android, Windows and Linux, and professional analysts went along with an AI agent’s secretly assigned leanings 86% of the time. Artificial Analysis launched Intelligence Index v5, built mostly on private test sets, and the gap to open-weights models widened sharply; on compute, Google TPUs are leaving Google for the first time to land on Oracle Cloud, and media reports say OpenAI plans to buy Google AI chips. In math, Axiom used AI agents to fully formalize the classification of finite simple groups in Lean in 24 days.

Today’s highlights

• Google DeepMind released Genie 4 with native audio and real-time natural-language scene edits; it opens first to trusted testers and is coming soon to Google AI Ultra and the Gemini API.
• Anthropic disclosed its acquisition of one-person biotech Mithril and published two studies: an unreleased model built five exploit chains (62 CVEs fixed as of Chrome 156), and analysts followed the agent’s leaning 86% of the time.
• Artificial Analysis Intelligence Index v5 is out: 60% is based on private or held-out evaluations, Claude Opus 5.5 (max) leads at 58, and the open-weights gap grew from 3 points to 12.
• Google TPUs will arrive on Oracle Cloud in 2027, the first time they appear on a hyperscale cloud outside Google; The Information reports OpenAI plans to buy Google AI chips [Media report].
• Axiom used AI agents to write about 43.1 million lines of Lean in 24 days, completing a full formalization of the classification of finite simple groups.

1. Google DeepMind releases the Genie 4 world model: native audio, real-time natural-language scene edits, and its first trip to the API

Launch: Google DeepMind released Genie 4 early on Oct 28 (Beijing time), calling it “our most realistic and interactive world model yet”: dynamic environments can be explored in real time, and people in the scene move, react and talk (21.2K likes, ~2.8M views). Sundar Pichai then listed the highlights: real-time control of events, camera and actions; native audio matched to the world (including characters you can talk to in real time); photorealistic scenes and physics; and more coherent multi-step events using Gemini reasoning (17.1K likes). Demis Hassabis reposted it, saying it combines real-time physics, reasoning and audio (3,657 likes).

Tech and limits: According to the official blog, appearance and behavior are built on Gemini Omni, with other Gemini models handling reasoning; native audio covers music, ambient sound and dialogue; and scenes can be changed in real time with natural language. Google acknowledges that object permanence drifts over time and that continuous interaction lasts a few minutes. The Verge says it uses Gemini 3.8 Flash and that a single session runs about 2 minutes versus about 1 minute for Genie 3 (our calculation: about 2x) [Media report].

Availability: It opens first as a research preview for trusted testers; it is “coming soon” to Project Genie for Google AI Ultra subscribers and to early partners on the Gemini API / Google AI Studio, with a waitlist open; generated audio and video carry SynthID watermarks. Logan Kilpatrick said this is the first time Genie comes to the API (2,485 likes). Ars Technica says API pricing has not been announced.

Sources: @GoogleDeepMind · @sundarpichai · @GoogleAI · @OfficialLoganK · @demishassabis · DeepMind blog · Google blog · The Verge · Ars Technica

2. Anthropic shares early drug-development progress and discloses its acquisition of “one-person biotech” Mithril Medicine

Official: Anthropic said that over the past few months it has worked closely with six drug-discovery organizations and has started its own drug-discovery work (8,870 likes, ~2.28M views). The partners named in the blog are BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines and The Column Group.

Mithril: Mithril Medicine was founded by Cody Tranbarger alone. Using Claude to coordinate 11 outside labs, he took a muscular-dystrophy antibody therapy from idea to early positive results in mice in under 18 months; Anthropic says the earliest human trials would be in late 2028 and the work is still mostly preclinical. Tranbarger (@Prof_Oak_) revealed himself in a post: “Anthropic acquired a one-person biotech . . . that person is me” (10.3K likes, ~940K views), with a technical report attached.

Reactions: @AndrewCurran_ said this fulfills the “one-person company” prediction (1,614 likes) [Opinion].

Details: R&D World reports the acquisition closed in July with undisclosed terms; the mouse study had 7 treated animals versus 4 controls, grip strength was about 2x the controls, and assignment was not randomized; the team also has 8 part-time advisers, and a second program targets 2029 [Media report]. The technical report has not been peer-reviewed [Not peer-reviewed].

Sources: @AnthropicAI 1 · @Prof_Oak_ 1 · @Prof_Oak_ 2 (444 likes) · @AndrewCurran_ · Anthropic blog · R&D World · TNW

3. Two Anthropic studies: an unreleased model builds five exploit chains; analysts follow an AI agent’s “leaning” 86% of the time

Exploit chains: According to Anthropic’s Frontier Red Team blog, an unreleased internal research model built five working exploit chains: (1) one-click remote code execution in Chrome with a sandbox escape (involving V8, a CacheStorage race and a FedCM use-after-free); (2) a one-click same-origin policy bypass in Safari (fixed by Apple); (3) an Android lock-screen bypass on Pixel 9 (fixed by Google); (4) a remote denial of service in Windows HTTP/3 (QUIC), CVE-2026-33096; and (5) a zero-click root remote code execution via Linux MPTCP, CVE-2026-98432. Anthropic says that as of Chrome 156 the work has led to 62 assigned and fixed CVEs, all critical or high severity, including 13 sandbox escapes and 21 renderer bugs. The Hacker News reported that one Chrome exploit pops a calculator on a Mac just by opening a web page (152 likes).

Expert-judgment study: Anthropic had 50 professional data analysts use Claude Code (Opus 4.8) for a causal-inference analysis, and on 4 analytic choices a coin flip secretly assigned the agent a “leaning.” Results: analysts chose the agent’s preferred option 86% of the time, versus about 50% with no influence (our calculation: 36 percentage points higher); analysts who had already stated a view changed their minds 60% of the time when the agent disagreed, versus 6% when it agreed (our calculation: 10x, but on only 38 cases, so the confidence interval is wide); afterward, 77% repeated the agent’s arguments and only 15% credited the source. Anthropic also notes that agent users were faster and made fewer serious errors (3 of 50, versus 6 of 11 without the agent), but the two groups are not directly comparable, and the experiment deliberately amplified the leanings, so real-world effects may be smaller.

Reactions: Co-author Alex Tamkin introduced the study (2,831 likes, ~416K views); Ethan Mollick called it an “important, but not unexpected, finding” and suggested deliberately planning when to use AI if independent judgment matters (461 likes) [Opinion]; @cremieuxrecueil highlighted the 86% and 77%/15% figures (862 likes). Al Jazeera also covered it.

Sources: @AlexTamkin 1 · @AlexTamkin 2 (229 likes) · @emollick 1 · @cremieuxrecueil · @TheHackersNews · Anthropic: Discovering vulnerabilities with Claude · Anthropic: How AI agents shape expert judgment · Al Jazeera

4. Artificial Analysis releases Intelligence Index v5: 60% private test sets, Opus 5.5 leads at 58, and the open-weights gap widens to 12 points

Launch: A day after posting only “Tomorrow.”, Artificial Analysis released Intelligence Index v5, adding 2 benchmarks and upgrading 3, and calling it the biggest update since v4 in January; 60% of the index is now based on private test sets or held-out-answer evaluations (4,528 likes, ~1.07M views). Four categories each carry 25%: knowledge work (AA-Briefcase, AutomationBench-AA, GDP.pdf), coding (FrontierCode 2.0, Terminal-Bench 4.0), science (Terminal-Bench-Science, the CritPt v2 hard subset) and general (AA-Omniscience v1.1, HLE-Diamond). It also launched Progress Index v1.

Rankings: Claude Opus 5.5 (max) 58, GPT-6 Astra (max) 57, GPT-6.1 Sol (max) 55, Gemini 4 Argon (high) 50; the top five are all Anthropic and OpenAI models. Cost to run the full index: about $8,910 for Opus 5.5 and about $5,661 for GPT-6 Astra (our calculation: about 1.6x) [Self-reported by the leaderboard].

Open weights: The leader, Kimi K3.1 Pro (weights not yet released), scores 46, 12 points behind first place, versus a 3-point gap in v4.3 (our calculation: the gap is 4x larger). Among models with released weights, DeepSeek V4.1 Flash 1017 (max) and GLM-5.3 (max) tie at 32, and Claude Haiku 5.5 (max) beats all of them. On HLE-Diamond and Terminal-Bench-Science, Kimi K3.1 Pro scores 37% and 34%, versus 68% for Opus 5.5 and 63% for GPT-6 Astra [Self-reported by the leaderboard]. Ethan Mollick said to “take with a grain of salt,” but it matches the general vibes that “we haven’t seen a Fable-class open model” (1,244 likes) [Opinion].

FrontierCode 2.0: Cognition released a new coding benchmark that asks not only “can this PR be merged” but “should it be merged,” with more varied, larger-scope tasks (2,727 likes); the launch post says the top model scores only about 45%, while AA measured Opus 5.5 at about 46%.

Rumor: @kimmonismus said a hidden model, “Obsidian 3,” shows up in Cognition’s data at 46.8%, above Opus 5.5’s 45.5%, and guessed from its use of the Claude Code harness that it is Fable 5.5 (633 likes); neither Anthropic nor Cognition has confirmed this [Unverified].

Sources: @ArtificialAnlys 1 · @ArtificialAnlys 2 (437 likes) · @ArtificialAnlys 3 (105 likes) · @cognition · @emollick 2 · @kimmonismus 1 · Artificial Analysis Intelligence Index v5 · Opus 5.5 vs GPT-6 Astra comparison

5. Google TPUs leave Google: Oracle Cloud to offer TPUs in 2027; The Information says OpenAI plans to buy Google AI chips

Oracle: On Oct 27 (US time, at Oracle AI World in Las Vegas), Oracle and Google announced that Google TPUs will come to Oracle Cloud Infrastructure (OCI), the first time TPUs appear on a hyperscale cloud outside Google. Customers can choose eighth-generation TPU 8t (training) and TPU 8i (inference), available from 2027 and scaling over multiple years. Oracle Cloud announced it on X (673 likes); SemiAnalysis joked: “PoV of Jensen after reading TPUs are coming to Oracle Cloud” (1,162 likes) [Joke].

OpenAI: @DeItaone relayed Reuters: citing people familiar, The Information reports OpenAI plans to buy Google’s AI chips; deal value and size were not disclosed, and Reuters could not independently verify the report (814 likes). The Reuters piece says OpenAI declined to comment and has been renting Google chips since June 2025 [Media report].

Sources: @OracleCloud · @SemiAnalysis_ · @DeItaone · Oracle press release · Google Cloud blog · The Information · Reuters (via Investing.com)

6. ChatGPT: iCloud Mail and Calendar plugin launches; halved limits for legacy Pro $200 users from Oct 30 spark complaints

iCloud plugin: ChatGPT announced that the iCloud Mail and Calendar plugin is live, letting users find emails, triage the inbox, and view and edit their schedule on web, desktop and iOS across all plans (5,421 likes, ~1.53M views). OpenAI’s Tibo (@thsottiaux), in a “Day 23” post, first said it was the iCloud plugin, then added “Just kidding, today is a reset” (7,538 likes, ~1.07M views) [Joke].

Pro limits: OpenAI’s help page says Pro $200 users who had kept legacy limits will keep them through Oct 29, after which they move to lower limits at the same price; a separate Pro 500 tier includes Ultrafast. The official page does not give the size of the cut; media and users put it at 20x Plus dropping to 10x, GPT-6 Pro weekly messages dropping from 200 to 100, plus a one-time grant of 62,500 credits (about $2,500, expiring Dec 31) [Media report].

Reactions: @zachmoskow said Codex, Work and GPT-6 Pro messages are all cut 50% while the price stays $200 (1,181 likes); @TokenGremlin said a one-time credit doesn’t make up for a permanent cut (408 likes) [Opinion].

Sources: @ChatGPT · @thsottiaux · @zachmoskow · @TokenGremlin · OpenAI Help Center: ChatGPT Pro tiers · MacRumors (iCloud plugin) · MakeUseOf (62,500 credits) · enjoy.dev (Oct 30 change)

7. Axiom uses AI agents to fully formalize the classification of finite simple groups in Lean in 24 days

Official: Axiom announced a complete automated Lean formalization of the classification of finite simple groups (CFSG, “the periodic table of symmetry”), calling it the largest mathematical proof in human history: the original runs about 15,000 pages and took more than 100 mathematicians across generations (6,547 likes, ~3.39M views). Per the official blog: about 43.1 million lines of Lean, written in 24 days by a swarm of AI agents directed by engineers and mathematicians (our calculation: about 1.8 million lines per day on average); the original proof spans more than 500 papers; the effort found 34 notable discrepancies with the published literature and produced one new result (Theorem 2.3); advisers include Lyons and Solomon; and the code is open source.

Reactions: Team member @andrew_v10209 cited more than 43 million lines of Lean, 715 papers and books, and 100 substantive corrections to the literature (1,750 likes), a different count from the official “34 notable discrepancies,” so we go with the official blog [One-sided claim]; @kimmonismus called it a milestone for AI handling an entire interconnected body of mathematics (993 likes) [Opinion].

Sources: @axiommathai · @andrew_v10209 · @kimmonismus 2 · Axiom blog · GitHub: AxiomMath/CFSG

In brief

• Trump says he will “ban” the term AI: On Oct 27 (US Eastern time), Trump posted on Truth Social that he will formally ban the term “AI/artificial intelligence” in all relevant documents “due to its inaccuracy” (@atrupar, 1,075 likes). Background: Executive Order 14434 of Sept 29 already told the executive branch to use “superintelligence (SI)” in non-statutory documents, without requiring existing documents to change. Conservative commentator Matt Walsh said he doesn’t know why the renaming matters so much to Trump (@MattWalshBlog, 20.8K likes, ~1.75M views) [Opinion]. Coverage: Reuters (via Investing.com) · The Hill · Variety · Nextgov (executive order background)
• Anthropic signs a OneGov deal with GSA: Claude Enterprise, Claude for Government and the Claude API are 40% off for government, extended for the first time to state, local, tribal and territorial governments, with no minimum spend and pay-as-you-go pricing; Anthropic says about 500 US government agencies use Claude and expects $1B in government savings [Company claim] (@AnthropicAI 2, 1,791 likes). Official: Anthropic · GSA press release
• AI science follow-ups: On OpenAI’s recently announced Yang–Mills result, KITP issued a statement from a “Yang–Mills working group”: a scientific collaboration led by David Gross will interpret and verify the AI result (@KITP_UCSB, 931 likes, reposted by OpenAI). On the claim that an Anthropic internal model refuted the Exponential Time Hypothesis (ETH), which had circulated on X and was previously labeled [Unverified]: Anthropic has now published on GitHub the paper “A 2^o(n) Time Algorithm for 3-SAT” (authored by Claude) plus an AI-written Lean 4 formalization, saying the result is purely asymptotic and irrelevant to practical solving; the README also notes that Lean only checks the proof, and whether the statement files match the paper needs human review. No independent peer review has appeared so far [Not peer-reviewed]. Sources: KITP statement · The Verge · anthropics/formal-math (eth)
• Anthropic IPO update: Bloomberg reports Anthropic picked JPMorgan to lead the US retail portion of its IPO and is considering allocating 15%–20% of shares to individual investors (@Cointelegraph, 241 likes). CNBC noted the S-1 shows $112.4B in cash and $19.3B in short-term investments as of June 30 (our calculation: $131.7B combined) (@CNBC, 56 likes). Coverage: Bloomberg Law · CNBC [Media report]
• Claude Duet widens access: @ahuja_karan said all US Max users are being prioritized off the waitlist, US Pro users will be invited next, and there is no date yet outside the US (@ahuja_karan, 1,046 likes) [One-sided claim].
• Liquid AI releases LFM3: The new generation of Liquid foundation models uses a tri-hybrid architecture plus a new multimodal encoder, with 6 models spanning text, audio, vision and omni-modal; LFM3-1.7B supports 256k context (@liquidai, 7,662 likes). Official: Liquid AI blog
• Cursor design mode adds Annotate: Edit elements on the page, draw on it, and check designs across screen sizes to give the agent visual prompts (@cursor_ai, 3,618 likes).
• Gemini can check for SynthID: Ask Gemini “Is this AI-generated?” to check whether an image, video or audio clip was made with Google AI (@GeminiApp, 1,197 likes).
• Microsoft MAI-Image-2.7-Preview: Ranks 2nd on AA-Image-Editing v2.1, behind only OpenAI GPT Image 2.5 Sunburst (max), and beats MAI-Image-2.6 on all 7 editing actions (@ArtificialAnlys 4, 470 likes) [Self-reported by the leaderboard].
• Inference speed and funding: AMD says DeepSeek V4.1 Flash hits 2,100+ tokens per second per user on 8 MI355X GPUs (implemented by lithos_ai) (@AIatAMD, 455 likes) [Company claim]; inference company General Compute announced a $100M Series A co-led by Blackbird and Abstract (@general_compute, 1,265 likes) [Company claim].
• Grok Bot 0.77.0: @ericzakariasson said the 0.77.0 stable release is rolling out, adding permission settings for auto-review rules, a preferred payment method option, and multiple questions in one card (@ericzakariasson, 1,212 likes); Elon Musk urged people to hire Grok Bot, saying they won’t regret it (@elonmusk, 8,126 likes, ~15.22M views) [Company claim].

---

About this column and how we verify

1. AI-assisted editing: AI Daily Brief is AI-assisted; this issue was produced with help from Grok Bot and covers Beijing time Oct 27, 2026, 22:17 through Oct 28, 2026, 22:15.
2. Tweet-level checks: Every cited post was fetched to confirm it exists, the author matches, and the timestamp falls in this period; all selected posts were also re-read via the X API. Like and view counts are as of editing time.
3. X News summaries are leads only: X News blurbs are Grok-generated and used only to find stories; facts are checked against original posts, official pages and mainstream coverage. When a summary conflicts with a primary source, the primary source wins. Figures marked “our calculation” were computed by us from public data.
4. Items about Grok Bot: The “Grok Bot 0.77.0” item in brief concerns the product used to help produce this column; it was included under the same standards as every other item.
5. Label key: [Unverified] = circulating without an official or other primary confirmation; [Media report] = from press coverage that may differ in framing from the official release; [Company claim] = from a company release or official account, not independently reproduced; [Self-reported by the leaderboard] = from a benchmark publisher’s own write-up; [One-sided claim] = from one party only, with no other source corroborating; [Not peer-reviewed] = a paper or technical report not yet peer-reviewed; [Opinion] = judgment, not a factual claim; [Joke] = said in jest by the poster, not a formal announcement.