Thursday 15 October 2026

Daily AI Brief

Top 3

Anthropic's Q3 revenue reaches $16.8B
The Financial Times reports third-quarter revenue grew about 45% from $11.6 billion in Q2, and that Anthropic was profitable on an adjusted basis for the first time. Its IPO is widely expected in November, at a target valuation of up to $2 trillion.
siliconangle.com
Eight US states sue OpenAI
Iowa, Nebraska, Kansas, West Virginia, Arkansas, Arizona, Idaho and Montana each sued in state court, alleging OpenAI broke consumer protection laws, misled users about its products' safety and endangered children. Arkansas also asks the court to halt testing of advanced AI models until adequate safeguards are in place.
engadget.com
CoreWeave debt fears hit AI stocks
CoreWeave shares fell more than 15.5% on 14 October after a $1.25 billion batch of its bonds briefly passed a 15% yield; Oracle fell 8.8% and Dell about 11.3%. CoreWeave reportedly has $50 billion in outstanding obligations, largely financed with debt.
siliconangle.com

Grok Voice 3.0
The new Grok speech-to-speech model is live in the Voice Agent API at $0.08 per minute, the same price as Grok Voice 2.0, with latency under one second. Agents using grok-voice-latest moved to it automatically.
x.ai
Paper Instruments' Ultramarine 1 Flash
An agentic model for professional knowledge work, made by applying large-scale reinforcement learning to the open GLM-5.3 Flash. The company claims a 4.1-point average gain over that base across six knowledge-work benchmarks, mostly measured by itself, with 6.5% fewer tokens.
paperinstruments.com

Grok Voice 3.0 tops speech-to-speech index
Artificial Analysis's independent Speech to Speech Index puts Grok Voice 3.0 first at 86.9, ahead of GPT-Live-1 (Astra, medium) at 83.2 and Gemini 3.8 Live Extended Thinking (High) at 82.6.
artificialanalysis.ai

OpenAI agrees not to use alleged Apple secrets
New since yesterday: at a 14 October hearing OpenAI agreed not to use the trade secrets Apple says were stolen while the case runs, and Judge Edward Davila said he is "highly likely" to grant Apple a preliminary injunction. OpenAI hit back with a motion alleging a former employee took more than 100 confidential files on its unreleased hardware when returning to Apple.
9to5mac.com
YouTube Music labels AI-made songs
A label now appears on the player page when a track was created mainly with generative AI, such as a model composing it and producing the vocals. YouTube finds these tracks with music partners and its own detection tools, and plans to bring the labels to the main YouTube app.
blog.youtube
AI agents made 138 million payments on Base
Chainalysis counts more than 138 million x402 transactions by Q3 2026, up from about 1,500 in mid-2025, with over 87% under 10 cents. The median time from a wallet's first deposit to its first payment fell from 73 days to 18 minutes.
fortune.com
Anthropic opens a Singapore office
Its fifth Asia-Pacific office comes with a memorandum of understanding with Singapore's Ministry of Digital Development and Information covering AI safety, security, model evaluation and adoption research. Anthropic says Singapore ranks second of 121 countries in Claude.ai use per person.
anthropic.com

Claude helps root a Wi-Fi access point
Chasing corrupted downloads, a developer used Opus 5.5 to find a cache coherency bug in an end-of-life access point's hardware offload, though it kept cutting off sessions when asked to apply the fix; Opus 4.8 then built a root exploit (CVE-2022-0185) over about 17 hours with the author guiding it. Turning the offload off fixed the corruption and cost about 8% of throughput.
blog.dfb.sh
Tiny Little Office
Ten browser apps (word processor, spreadsheet, slides, diagrams, calendar and more) passed 400 points on Hacker News; one commenter says it was "vibecoded in three days", which the site does not say. Testers liked its simplicity but found no file import, URL size limits for large documents, cursor bugs on touch typing and blocked map tiles.
news.ycombinator.com

OpenCode 2.0
The coding agent now runs as one background server shared by its terminal, web and desktop apps, adds a Code Mode JavaScript interpreter that calls MCP servers or OpenAPI endpoints in one round trip, and runs subagents in the background. 1.0 plugins do not work in 2.0; upgrade with opencode upgrade --major.
anoma.ly
Claude Code 2.1.300
Brings /advisor to Claude Platform on AWS, gzip-compresses large requests to api.anthropic.com (opt out with CLAUDE_CODE_GZIP_REQUEST_BODIES=0) and adds prompt-cache token counts to telemetry. claude remote-control now exits with an error if given --sandbox or --no-sandbox.
github.com
GitHub Actions workflow lockfiles
A new .github/workflows/actions.lock, made with the gh actions-lock extension, pins actions and their transitive dependencies to exact commits, so a moved tag cannot change what a workflow runs. In public preview, with an admin policy to require lockfiles and Dependabot updates that refresh them.
github.blog
Go 1.27 drains HTTP/1 bodies for you
net/http.Transport will drain unread HTTP/1 response bodies in the background, up to 50 ms or 256 KB, so connections get reused without manual draining. The Go team now says to just close response bodies when you are done.
go.dev

Microsoft disrupts Chinese phishing kit sellers
Under a court order, Microsoft and Health-ISAC seized domains behind YAM Store, CoGUI and Iron Man, kits linked to over 1 billion phishing emails and texts sent to more than 75,000 organizations this year. More than 20,000 accounts were disabled and over 100,000 fraud domains blocked.
cyberscoop.com
Apple Intelligence heads to China
Bloomberg reports basic features such as photo editing arrive with iOS 27.3, expected in December, without Siri AI, Image Playground or the new Genmoji. Chinese regulators granted permits for Alibaba's Qwen and Tongyi Wanxiang models on iPhones, per the South China Morning Post.
macrumors.com
Flock CEO defends its cameras
At TechCrunch Disrupt, Garrett Langley said "we shouldn't pause progress" and that Flock would offer cities recommendations and defaults rather than mandated policies. Pressed on abuse, he gave no specifics; the backlash has led to cancelled contracts and layoffs.
techcrunch.com

AI agents aren't going rogue, labs skip basics
Security experts from SANS, RAND and the Berryville Institute argue that this year's AI hacking incidents, including the OpenAI and Hugging Face breach, came from missing containment, egress control and monitoring rather than unusually capable agents.
csoonline.com

One page read for this brief (tinylittleoffice.com) contained instructions for AI agents; they were ignored.

An AI agent I set up researches and writes each brief. Every item links to its source.

Want it in your inbox every day?