The Frontier AI Wire is researched and drafted by Claude, an AI model made by Anthropic, under rules set by Attorney Jeffrey M. Beck. Every factual claim links to its source, with primary sources first. Where the brief goes beyond what a source says, it labels that as inference. Attorney Beck reviews and approves each edition before it is published. Errors are corrected in place, marked where they occurred, and logged. Nothing is changed silently. How the Wire is made.
A frontier model went to the top of the independent leaderboard on a day when almost nobody could use it. Google announced Gemini 4 Carbon, which Artificial Analysis scores at 59, one point above Claude Opus 5.5, while access is limited to vetted cyber defenders and no model card was found. The same week's other cyber release sat unread for a day: Anthropic's red team says an unreleased model built working exploit chains against Chrome, Safari, Android, Windows and Linux, all now patched. Cloudflare published how a configuration mismatch and an inference-engine bug sent 155 customers' long-prompt responses built partly from other customers' cached data. No new corrections; two items from earlier windows run today with their dates.
Dispatches
Ranked by how much each item should change your picture of the field — not by volume of coverage.
Gemini 4 Carbon takes first place on Artificial Analysis's index by one point, and ships first to cyber defenders only
Google announced Gemini 4 Carbon on 28 October at 17:00 UTC, in a post by Koray Kavukcuoglu, two weeks after Gemini 4 Argon became generally available (edition 024). Google calls it its “next frontier model” and leads on long-running agentic coding and cybersecurity. Access starts with a subset of cyber defenders and widens through the Fairwind Program; developers, businesses and consumers get it “as soon as possible,” with no date. Introductory pricing matches Argon at $2 / $10 per million input and output tokens, rising to $4 / $20; Google does not say when the introductory period ends. Nathan Lambert, on the Frontier Wire Sources list: “Don't bother announcing your scores if we can't use it.”
Two independent evaluators ran it on launch day. On Artificial Analysis's Intelligence Index v5, Carbon (high) scores 59, first of 52 models, against 58 for Claude Opus 5.5 and 57 for GPT‑6 Astra (edition 034). On AA's Coding Agent Index, AA places it second, behind Claude Code with Opus 5.5. Vals AI ran it on Vals Index v2.2 at 70.36%, second of 29; Vals's own leaderboard lists Claude Sonnet 5.5 first at 70.85%. VentureBeat's account of Google's comparison chart sets Carbon's Vals score against Opus 5.5's 70.2% only.
| Measure | Carbon | Comparison | Source |
|---|---|---|---|
| AA Intelligence Index v5 (high) | 59 | #1 of 52; Opus 5.5 58, Astra 57 | Artificial Analysis |
| AA cost per index task | $7.76 | 200M output tokens to run the index; median 64M | Artificial Analysis |
| AA Coding Agent Index | #2 | Behind Claude Code with Opus 5.5 | Artificial Analysis |
| Vals Index v2.2 | 70.36% | #2 of 29; Sonnet 5.5 70.85%, Opus 5.5 70.18% | Vals AI |
| Terminal‑Bench 4.0 | 63.13% | #4 of 57 | Vals AI |
| SEC‑bench Pro | 88.2% | Argon 83.4% | |
| ExploitGym | 54.8% | Argon 42.0% | |
| CWE‑bench | 71.0% | Argon 68.0% | |
| WebDev Arena | 1768 Elo | #4, up 89 points | Google, citing the arena |
Artificial Analysis and Vals AI are independent companies that run the same tests on every new model, so labs cannot simply grade themselves. Carbon now has the highest score on Artificial Analysis's main index, by a single point — a margin small enough that “roughly level with the best Anthropic and OpenAI models” is the safer reading than “clearly ahead.” On Vals AI's index it comes second.
What it does not mean: that you can use it. For now it goes to a vetted group of security organisations, and Google has not said when that widens. The brief's read: Gemini 4 Argon also went first to cyber defenders (edition 015), so for Google, security capability now sets the release order. The premise is that the Fairwind Program is a safety gate rather than a sales channel; Google's post describes refining guardrails before broad release, which supports it, and says no more.
What Google reports about behaviour, all its own measurements. Successful cheating on Terminal‑Bench 4.0 at 0.0%, with attempts at 0.4%; 0.0% attack success on Gray Swan's indirect prompt-injection benchmark against Argon's 0.7%; unsupported claims roughly halved against Argon; overreach down 26% and under-service down 29%; and high-severity misalignment alerts in internal coding environments down almost 80%. Google says its cyber benchmarks were run without guardrails, that Wiz observes Carbon leading on web penetration-testing benchmarks, and that the Chrome team saw it produce high-quality patches “more than 10X better” than an unnamed lab's large model. Google also says it is taking part in the US government's voluntary pre-release testing.
Where the independent numbers cut the other way. VentureBeat, reporting Artificial Analysis's component results, says Carbon leads AA‑Omniscience at 54 but that its hallucination rate rose from 4% to 11% against Argon, and that it ranks below Grok 4.7, Muse Wonder and Kimi K3.1 Pro on AA‑Briefcase. Those component figures were not on AA's model page as read at compile time; they are VentureBeat's account of AA's run.
What is missing. No Gemini 4 Carbon model card was found at compile time on DeepMind's model-card pages, whose latest-cards list still shows Gemini 4 Argon. Google's post states no context window; AA's page lists 1M tokens. AA's page lists the post-introductory $4 / $20 rate and is internally inconsistent on whether a public API exists. No architecture or parameter count is published.
Sources Google, 28 Oct (primary) · Artificial Analysis model page · @ArtificialAnlys — read 29 Oct from the Frontier Wire Sources list · Vals AI model page · Vals Index leaderboard · VentureBeat, 28 Oct · @natolambert
Anthropic's red team says an unreleased model built working exploit chains against five platforms, all reported and patched
“Discovering vulnerabilities with Claude,” dated 27 October, inside edition 034's window and not carried there. Anthropic's Frontier Red Team says “an unreleased, internal research model,” which it does not name, built the exploits below “after initial prompting,” with “little human involvement,” over “the past few months.” The post gives no cost figures and does not say whether the model will be released.
| Target | What the exploit did | Status |
|---|---|---|
| Chrome (macOS) | One-click remote code execution: a four-stage chain from V8 heap corruption through a sandbox escape | Patched in Chrome 156; Anthropic says Google assigned 62 CVEs, all critical or high |
| Safari | One-click same-origin-policy bypass reading history, cookies and session data | Fix available; no CVE given |
| Android (Pixel 9) | Photos and personal data extracted from a locked phone | Addressed by Google; no CVE given |
| Windows | Remote crash of machines answering HTTP/3 requests | CVE‑2026‑33096 |
| Linux | Zero-click root code execution via a multipath TCP race; 48 minutes to leak a kernel pointer, 5 more to a shell | CVE‑2026‑98432, patched upstream and backported |
An “exploit chain” strings several separate software flaws together so that, for example, visiting one web page lets an attacker run their own code on your computer. Building one against a hardened browser or operating system has traditionally taken skilled specialists weeks or months. Anthropic says one of its unreleased models did this against five of the most widely used platforms, and that every flaw was reported to the vendor and fixed before publication.
What it does not mean: that these attacks are usable now — the post says they are patched — or that any public Claude model can do this. The brief's read: taken with item 01, two labs in two days led with offensive-security capability, one by releasing a model to defenders first and one by describing a model it has not released.
What the public records show. The Hacker News checked the vendor records and found them less specific than Anthropic's account: the CVE‑2026‑33096 record names HTTP.sys, credits no finder and is marked not exploited; the CVE‑2026‑98432 record describes an IPv4 TCP fix rather than a multipath TCP one; Chrome's release notes credit an Anthropic researcher “assisted by Claude” for some, not all, of the fixes; and Apple's iOS 26.5.2 notes credit Anthropic researchers “with Claude.” None of that contradicts the post. It is the extent of outside confirmation found at compile time.
Disclosure terms and access. Anthropic says it followed its coordinated disclosure policy: 90 days for vendors to patch, then 45 more before discussing details beyond a public advisory. The post also expands the Cyber Verification Program (edition 019), which gives vetted security professionals Opus 5.5, Sonnet 5.5 and Mythos 5.1 with reduced blocking classifiers, and releases a Claude Code security-guidance plugin. It describes no safeguards for the internal model itself.
Sources Anthropic Frontier Red Team, 27 Oct (primary) · The Hacker News, 28 Oct (record checks) · CVE‑2026‑33096 · CVE‑2026‑98432 · Apple, iOS 26.5.2 notes
Cloudflare: an undersized cache and an inference-engine bug sent some customers responses built partly on other customers' data
Cloudflare's post, dated 26 October and published jointly with the SGLang maintainers, inside edition 033's window and not carried there. Between 27 August and 9 September, 155 responses from the GLM‑5.3 model on its Workers AI service were generated partly from data left by other customers' earlier requests. Every affected prompt was longer than 681,280 tokens. A researcher at Knostic reported it on 8 September through HackerOne; Cloudflare took the model private the next day, fixed it and restored it the same evening. It found no evidence of deliberate exploitation and says no other Workers AI model was affected.
To answer quickly, an AI service keeps a working memory of each prompt it is processing, called a cache. Here, the cache was smaller than the longest prompts the service accepted. When a prompt did not fit, a software bug let the request carry on anyway, and the model read leftover cache contents from someone else's earlier request. A customer could receive an answer influenced by another customer's text.
Why it matters: AI services are shared machines, and this is a documented case of one customer's data reaching another through the model-serving layer rather than through a conventional breach. The brief's read: it is the kind of failure a provider's tests miss — Cloudflare says its tests checked that requests succeeded, not what came back.
The two faults. The model was deployed with a 1,048,576-token context window but a prefill KV pool of 681,280 tokens; Cloudflare says it knew of the mismatch before launch and its tests checked HTTP status codes, not output. In SGLang, an over-length prompt “was internally marked as rejected and replaced with a one-token placeholder,” but the admission queues checked whether a request was finished rather than rejected, so it went through the prefill-to-decode handoff and decode attended over stale KV pages. The same bug had been reported upstream on 6 July as issue #30233.
Fixes and the operator advice. The prefill pool now holds about 2.04 million tokens, roughly twice the context length; rejected requests are retired before any KV transfer (upstream fixes #36988 and #38935); deployment checks compare pool size to context length; and Cloudflare is working on per-page integrity tags. It recommends that “all operators update SGLang dependencies to v0.5.22 or later.” The post does not describe notifying the affected customers directly.
Sources Cloudflare, 26 Oct (primary) · The Neuron digest, 28 Oct (where it surfaced)
Microsoft says it booked $10 billion of revenue from OpenAI in one quarter
TechCrunch reported on 28 October that Microsoft's regulatory filing for the quarter to 30 September — the first of its fiscal 2027 — shows $10 billion of revenue from OpenAI under commercial arrangements including revenue sharing, nearly 11% of Microsoft's $92.3 billion for the quarter, against $24.1 billion for all of fiscal 2026. Receivables from OpenAI rose to $7.2 billion from $6 billion at the end of June, and Microsoft's stake fell to about 24% from about 25%; Microsoft did not say what caused the dilution. The split between revenue sharing and Azure is not disclosed, and TechCrunch says it has asked. The figures describe Microsoft's revenue; they say nothing about OpenAI's costs or margins, and this brief draws no conclusion about either.
Sources TechCrunch, 28 Oct
Also on the wire
Confirmed, and not enough on their own to move the ranking.
-
Anthropic extends AI for Science credits to foundations and institutes
On 28 October Anthropic said its AI for Science programme, which gives free Claude credits to researchers and began in May 2025, now runs through funders, institutions and named projects, among them the Sloan and Simons foundations, Schmidt Sciences, the US National Science Foundation, the Wellcome Sanger Institute, the Crick, EMBL, FlyWire and gnomAD. The headline says 20 partners; the body names 18. No credit amounts are given.
Source Anthropic (primary)
-
ARC Prize verifies Meta's Muse Spark Open: 57.5% on ARC‑AGI‑2, near zero on ARC‑AGI‑3
Posted 28 October: on the semi-private sets at xhigh reasoning, 89.0% on ARC‑AGI‑1 and 57.5% on ARC‑AGI‑2 ($0.73 a task), and 0.1% (standard harness) and 0.4% (provider adapter) on ARC‑AGI‑3. ARC Prize says its v1 and v2 scores are similar to Claude Haiku 5.5 and GPT‑6 Luna “but ~9x the cost.” The model shipped on 15 October (edition 026).
Sources ARC Prize results page · @arcprize — read 29 Oct from the Frontier Wire Sources list
-
Mistral launches a Responses API and starts serving DeepSeek and Z.ai models
On 28 October Mistral put a Responses API for building agents into public preview and said it now serves DeepSeek V4.1 Flash (public preview) and GLM 5.3 Flash (generally available) on its EU and US cloud. Mistral's docs list both as third-party, MIT-licensed models with 1M-token context.
Sources @MistralAI — read 29 Oct from the Frontier Wire Sources list · Mistral docs
-
Reuters: the Justice Department is asking about Nvidia's Hugging Face purchase
Reuters reported on 28 October, citing two people familiar with the matter, that DOJ officials are asking whether Nvidia's roughly $12.93 billion purchase of Hugging Face, agreed in early September and expected to close in the first half of 2027, would give it control of a key AI distribution channel. Reuters said the inquiry may not lead to action. Read here through PYMNTS's republication.
Source PYMNTS, citing Reuters
-
Alphabet's quarterly capital spending reaches $51.4 billion
Alphabet reported on 28 October third-quarter capital expenditure of $51.4 billion against $24 billion a year earlier, free cash flow of minus $9.1 billion, and Google Cloud revenue of $30.7 billion, per Fortune. See Checked and spiked on the Cloud growth rate.
Source Fortune, 28 Oct
-
DOE offers Constellation a $5 billion loan for nuclear uprates sold to Google, Amazon and Meta
Announced on 27 October and not carried in edition 034: a conditional commitment, not yet final, toward uprating 11 reactors at eight plants for about 1.1 GW. TechCrunch reports that Google, Amazon and Meta have bought most or all of the added output at seven of the eight plants.
Source TechCrunch, 28 Oct
-
OpenAI gives Pro and Business subscribers up to $500 a month of API credit
OpenAI's developer changelog for 28 October: eligible ChatGPT Pro and Business subscribers can claim up to $500 a month in free API credits by linking an API organisation they own.
-
The standing disclosure pages, read this morning
OpenAI's misalignment page still prints 20 of 20 reports and three Notices; the newest entry was first posted and last updated 20 October. The Deployment Safety Hub's newest item is still the 26 October GPT‑6.1 Terra addendum. Anthropic's coordinated vulnerability disclosure dashboard still serves its 2 October 19:47 UTC snapshot: 6,157 disclosed across 591 projects, 516 patched, 219 CVEs and 365 GitHub advisories. The Transparency Hub's Model Report is dated 2 October.
Sources OpenAI misalignment reports · OpenAI Deployment Safety Hub · Anthropic CVD dashboard · Anthropic Transparency Hub (all read 29 Oct)
-
The lab desks, as read this morning
anthropic.com/news's newest item was the 28 October AI for Science post; its research index's newest was 27 October (item 02 and edition 034's items). openai.com/news did not render at compile time; OpenAI's developer changelog carried one 28 October entry. mistral.ai/news's newest was 21 October; the Responses API ran on X. x.ai/news returned only August items as fetched, so nothing is asserted about it. ai.meta.com/blog's newest was still 27 July. DeepSeek's news link still pointed to its 17 October item. Qwen's Hugging Face organisation's newest model was about 16 days old. METR's newest was 22 October, Epoch AI's 27 October. DeepMind's index printed month-only dates; Carbon was dated from the post itself. These describe what those pages showed at compile time, not whether the organisations published elsewhere.
Sources anthropic.com/news · anthropic.com/research · openai.com/news · Mistral · x.ai/news · METR · Epoch AI · Qwen on Hugging Face
Checked and spiked
Items that circulated but did not survive verification.
“Google Cloud revenue doubled.” Circulating with Alphabet's results, including as 103% growth in one aggregator. Fortune's own account carries the “doubles” headline and, in its body, Sundar Pichai saying Cloud revenue grew 48%. Those cannot both describe the same year-on-year comparison, and Alphabet's filing could not be retrieved at compile time to settle it. The $30.7 billion figure is carried; no growth rate is.
Sources Fortune, 28 Oct · AI Weekly (the 103% figure)
Corrections
Errors in this brief — fixed in place above, logged here.
No new corrections. Editions 002, 003, 006, 008, 014, 015 and 025 carry corrections, archived with their editions. Recorded rather than corrected: Anthropic's vulnerability post (item 02, dated 27 October) and Cloudflare's KV-cache post (item 03, dated 26 October) fell inside the windows of editions 034 and 033 and were not carried there, and the DOE loan in Also on the wire was announced on 27 October; nothing those editions said was wrong. Edition 034's statement that Claude Opus 5.5 led Intelligence Index v5 at 58 was accurate when written and has been overtaken by item 01, which is not an error. The standing rule after edition 015 — that an absence is reported as “not found at compile time, at this URL” or not at all — applies throughout, including the model card in item 01 and the lab desks. On sourcing: the private source list was reachable at compile time through the in-app browser; the Artificial Analysis result for item 01, Lambert's reaction, the ARC Prize result and the Mistral announcement came through it. Chrome was not reachable.