K Koda Intelligence
Subscribe
Cover image for the 22 September 2026 Signal: Nikkei says AI cuts model development to a third; one HEIC upload cut into OpenAI employee accounts in July.
THE SIGNAL · 22 SEP 2026 · 12 MIN READ

Development time falls to one third

Nikkei reports Anthropic, DeepSeek and others are cutting frontier model development time to a third as AI takes over more of its own build cycle. Meanwhile one malformed HEIC upload reached inside OpenAI employee accounts in July.

KI Koda Editorial · Updated 01:10 GMT · 17 sources
AISECURITYGEOPOLITICS
26 of 34 figures cited14 receipts in the register
  1. 13%1Venturebeat
  2. 75x2Venturebeat
  3. $2.503Meta, via TLDR
The register
TL;DR
Today in three lines
  1. 01

    Nikkei says AI cuts model dev time to a third Nikkei reports Anthropic, DeepSeek and other labs are upgrading frontier models with development time cut to one-third...

  2. 02

    TypeSafe opens Jev judgment model to everyone TypeSafeAI opened Jev, a classification-only model from ex-OpenAI researcher Diogo Almeida, to all users, claiming 75x...

  3. 03

    US and China formalize an AI risk channel Treasury Secretary Scott Bessent and Vice Premier He Lifeng agreed to convert this week's New York pre-summit talks...

KODA PROFrom the desk

The Operator Tier is coming.

A weekly operator deep dive, the full Dojo Pro prompt packs, and the complete prompt database. Founding members lock the launch price forever.

Join the founding list →
Lead Story

The day's defining move.

Enterprise · Venturebeat
Lead storyEnterprise Enterprise Venturebeat · 22 September 2026
Enterprise·Venturebeat·18 SEP 2026

Researchers Detail July Takeover Of OpenAI Accounts

Researcher s1r1us disclosed a July 25 chain: a libheif heap overflow via HEIC upload gave RCE on community.openai.com, and an SSO flaw let them take over employee ChatGPT and Codex accounts and open an internal repo PR.

Why it mattersConnected GitHub, Slack and Outlook integrations turned one forum bug into repo access, so audit which third-party services your assistant accounts can reach and whether SSO scopes them.

Continue reading
World

Houthi fighters are moving on strategic heights that would cut Yemen's Red Sea coast off from Saudi-backed forces after Trump reportedly called off US strikes as bombs were being loaded, while Britain said RAF aircraft will refuel Saudi jets.

Markets

Tech shares drove the Nasdaq to a record close on Monday as Brent crude fell to a two-week low on bets that UN General Assembly diplomacy could ease the Iran and Yemen confrontations, with sentiment at Extreme Greed.

Wild Card

Nikkei reports labs cutting model development time to a third by letting AI build AI, and Anthropic now lets Claude Code threads keep working after you close your laptop; the OpenAI disclosure shows the human-run infrastructure around those systems can still fall to a single image file.

Listen & Watch

Daily broadcasts.

Podcast · Video · Infographic
YouTube · Short

Today's Signal Short

Visual

Intelligence map

Markets

6 levels at the close.

Market TerminalLAST CLOSE, READ 2026-09-22 · Yahoo Finance
S&P 500 7,764.708 +1.49%8 6D, window short 7,551.81 low / 7,764.70 high
Nasdaq 27,122.099 +2.26%9 6D, window short 25,978.43 low / 27,122.09 high
Bitcoin $86,425.6010 +6.51%10 6D, window short $76,150.32 low / $86,425.60 high
Ethereum $2,774.0311 +4.96%11 6D, window short $2,416.12 low / $2,774.03 high
Crude Oil (WTI) $92.2912 -7.99%12 6D, window short $92.29 low / $105.83 high
Crypto Fear & Greed 78 / 10013 Extreme Greed 0 to 100 index 25 fear / 50 neutral / 75 greed · alternative.me
Today's Focus

What else the day turns on.

01

Agents Outlast The Session

Anthropic's beta routes a single Claude Code conversation into parallel threads that continue after the user disconnects, available to select Pro and Max users in cloud sessions. Meta's Muse, marketed as taking tasks off users' plates, topped mobile app charts this week per Sina Finance. Both point the same direction: agents that act without a human watching, which raises the value of the guardrails and access controls the OpenAI incident exposed.

02

Washington And Beijing Talk Risk

Treasury Secretary Scott Bessent and Vice Premier He Lifeng agreed to turn this week's New York pre-summit talks into a standing bilateral mechanism on AI risk ahead of the Trump-Xi meeting. Nikkei's report that Anthropic, DeepSeek and others are compressing development cycles to one-third shows why the timing matters: the race the channel is meant to manage is accelerating as AI takes on more of its own development.

Deep Dive

Today's editorial.

All editorials
Deep Dive cover image: One Photo Cracked OpenAI's Code Vault
Deep Dive2 min read

One Photo Cracked OpenAI's Code Vault

Read full analysis
The Wire

9 stories from 6 desks.

Lead carried above
Agents·Venturebeat·17 SEP 2026

Claude Code Runs Whole Projects From One Chat

Anthropic said Projects now run from one Claude Code conversation, with Claude routing requests to parallel threads that keep working after you close your laptop. The beta covers select Pro and Max users in cloud sessions.

Why it mattersThreads that keep running unattended in cloud sessions will bill tokens while nobody is watching; set usage caps before letting a Max seat delegate a whole project.

Read
Model Release·Venturebeat·21 SEP 2026

TypeSafe Opens Jev Judgment Model To All

TypeSafeAI opened Jev, a classification-only model from ex-OpenAI researcher Diogo Almeida, to all users. It claims 75x2 speed and 1/170th the cost of OpenAI's model; 13%1 of Vercel paid customers tried it in a day.

Why it mattersIf the 1/170th cost claim holds on your routing, scoring and moderation workloads, a chunk of LLM spend is misallocated; the launch-week API outage means capacity still needs verifying.

Read
Agents·CNBC·21 SEP 2026

Meta's Muse Agent Tops Mobile App Charts

Meta's Muse, the agent it markets as taking tasks off users' plates, rose to the top of mobile app rankings this week, per Sina Finance, giving Meta its strongest showing yet in consumer AI assistants.

Why it mattersChart position measures downloads, not retention; anyone repricing Meta's consumer AI prospects should wait for daily active figures before treating Muse as a ChatGPT rival.

Read
Also on the wire
  1. Nikkei AsiaNikkei: AI Cuts Model Dev Time To A Third
  2. ReutersAikido Releases Open-Weight Security Model Altar
  3. ReutersOpenAI Urges US Lead On Global AI Standards
  4. OpenAIReport: OpenAI To Unveil Managed Agents At DevDay
  5. The Manila TimesMeshy 7.1 Hits 4096³ Geometry Resolution
  6. OpenAIOpenAI Forms Mathematics And AI Advisory Group
The Globe

4 stories off the AI beat.

Conflict·The Guardian·21 SEP 2026

Trump Halts Houthi Strikes; UK Backs Saudi Jets

Houthi fighters are moving to take strategic heights that would cut Yemen's Red Sea coast off from Saudi-backed forces after Trump reportedly called off US strikes as bombs were being loaded. Britain said it will give Saudi aircraft RAF aerial refuelling.

Why it mattersA Houthi hold on the Red Sea coast keeps Saudi crude routing east through the Gulf and Hormuz, sustaining the shipping and insurance premium on Middle East energy.

Read
Also on the globe
  1. ReutersUS, China Formalize AI Risk Consultation Channel
  2. ReutersNasdaq Hits Record As Oil Slides
  3. ReutersNorth Korea Tests New Combat Weapon System
The Lab

Tool of the day, field tested.

All Lab reports
Vapi home page, captured for the 22 September 2026 Lab report Tool of the day 7.014/ 10
Voice

Vapi

Ring pushed 100% of its inbound support calls through Vapi within two weeks; the Reddit builder who logged 760 calls says the AI was 20% of the work and the other 80% is still yours.

The plumbing is the product here: telephony, transcription, turn-taking, and call state arrive pre-wired, and you plug in whichever LLM, STT, and TTS vendors you want from a roster that includes OpenAI, Anthropic, Deepgram, Cartesia, and ElevenLabs. Adopting it is cheap on day one and expensive in ways the headline pricing does not show, because model, speech, and carrier costs stack on top, and community reports of bugs, latency spikes, and thin support mean production is a project, not a weekend. Worth it for engineering teams that want to own their voice stack without building it; skip it if you need a turnkey agent with a predictable invoice.

Capability 8.0
Ease 6.5
Value 5.5
Momentum 7.5

Fast and developer-friendly; you get voice AI without wiring up telephony, transcription, and call state by hand, and the API, Web SDK, docs, and support all hold up.Paraphrased · Product Hunt

Read the Lab report Try it
Also on the radar
Build

Composio: 1,0004+ ready-made tools for the agent you already built

Composio sits between your agent and the outside world, handling authentication, permissions, and sandboxed execution for more than 1,0004 production-ready tool integrations, so the agent can act instead of just plan. Pricing is contact-for-quote, which signals a business and enterprise focus rather than a hobbyist tier. Start by wiring one integration your agent currently fakes with a stub, then compare its error rate against your hand-rolled auth before you migrate the rest.

Try it
Build

Labelf AI: train a text classifier with active learning and no ML background

Labelf AI is a low-code platform that lets anyone train text classification models, using active learning to surface the examples that most improve the model so you label less data for the same accuracy. It accepts both labeled and unlabeled data and has free options to start. Bring one messy dataset you already sort by hand, such as support tickets or survey comments, label a few hundred rows, and measure how quickly the suggested-label accuracy climbs before committing more time.

Try it
Productivity

Ernest: finds lookalike prospects and drafts the first email in your voice

Ernest scans for companies that resemble your existing clients, then drafts tailored outreach emails matched to your writing style, delivered through a Gmail extension so you never leave the inbox. It is listed as free on Product Hunt. Feed it your five best current customers rather than your whole client list, so the lookalike search reflects who actually pays and renews, and read every draft before it goes out; the tool writes the opener, you own the claim.

Try it
Productivity

Presto: turn screenshots and clips into how-to guides that close tickets

Presto assembles text, screenshots, and video walkthroughs into organized how-to guides you can share instantly with customers, aimed squarely at resolving repeat support tickets and doubling as lightweight product demos. Pricing is not specified in the listing, so check the site before rollout. Start with your single most-repeated ticket, build one guide for it, and track whether that ticket volume drops over two weeks; that number decides whether the tool earns a place in your support stack.

Try it
The Arena

2 labs moved this week.

  1. OpenAIPushes Washington to lead global frontier AI standards while quietly building out math advisory and education arms01
  2. Meta AIAmazon blocks Meta's Muse shopping agent days after launch, exposing the platform-access problem for consumer agents02
Builder Radar

3 receipts from the repos.

Primary sources, pulled by index
Receipts
  1. Behind: Researchers Detail July Takeover Of OpenAI Accounts Issue openai/openai-cookbook#2833GPT-5.5 in Action: AI-Automated Discovery of CVSS 9.8 Enterprise Vulnerability2833

    GPT-5.5 in Action: AI-Automated Discovery of CVSS 9.8 Enterprise Vulnerability2833.Description. Exploit-Garbage Hi OpenAI team, We want to showcase how GPT-5.5 played a critical role in our AI-automated vulnerability discovery system that...

    github.com
  2. Behind: TypeSafe Opens Jev Judgment Model To All DocsTypeSafe AI models

    TypeSafe AI models.

    vercel.com
  3. Behind: Composio: 1,0004+ ready-made tools for the agent you already built Issue composiohq/composio#4141Summary. Composio's new pricing went live on Aug 15, 2026: Hobby ($0) / Pro ($29/mo) / Enterprise (custom).

    Summary. Composio's new pricing went live on Aug 15, 2026: Hobby ($0) / Pro ($29/mo) / Enterprise (custom). This PR updates the public docs to describe only the current offering and reprices premium tools as pass-through (provider cost +...

    github.com

Receipts come from the Firecrawl Developer Index (READMEs, docs, issues, merged PRs), first-party sources only. The paper comes from the Firecrawl Research Index, arXiv, last 7 days, picked for today's lead. Quotes are verbatim.

THE DOJO · BUILD TODAY

Harden the upload path, then let the agent run long.

  1. 01

    Audit your image pipeline today. The s1r1us chain started with a malformed HEIC hitting libheif and ended in remote code execution on OpenAI's community forum. Pin and patch your decoders, strip metadata, and decode untrusted images in a sandboxed worker rather than your app process.

  2. 02

    Test one long-running agent job. Anthropic's beta now runs a whole project from a single Claude Code conversation, routing requests to parallel threads that keep working after you disconnect. Pick one multi-file refactor, hand it off, close the laptop, and review the diff instead of babysitting the session.

  3. 03

    Swap a judgment call onto a cheap classifier. Route your moderation or routing decisions through a classification-only model like Jev, and wire voice or tool flows with Vapi and Composio so the expensive frontier model only sees the cases that need it. Measure cost per decision before and after.

The Bottom Line

Speed is compounding, and so is the attack surface

If frontier development time really is down to a third, the gap between labs will be decided by shipping discipline rather than by ideas. That same speed shows up in your stack: agents that outlive the session, classifiers at a fraction of the cost, and open-weight security models like Altar you can run locally. But the July OpenAI chain is the reminder that a single image parser can undo all of it. Move fast on capability, and treat every untrusted input as the thing that gets you.

Receipts

Every number, and where it came from.

26 of 34 figures cited, 14 receipts: 0 verified, 12 reported, 2 estimated
  1. 0113%It claims 75x speed and 1/170th the cost of OpenAI's model; 13% of Vercel paid customers tried it in a day.Venturebeat21 SEP 2026as publishedR
  2. 0275xIt claims 75x speed and 1/170th the cost of OpenAI's model; 13% of Vercel paid customers tried it in a day.Venturebeat21 SEP 2026as publishedR
  3. 03$2.50...offering text-prompted detection, segmentation, and tracking at $2.50 per 1,000 images or $0.20 per 1,000 video framesMeta, via TLDR22 SEP 2026as publishedR
  4. 041,000...text-prompted detection, segmentation, and tracking at $2.50 per 1,000 images or $0.20 per 1,000 video framesMeta, via TLDR22 SEP 2026as publishedR
  5. 05$0.20...segmentation, and tracking at $2.50 per 1,000 images or $0.20 per 1,000 video framesMeta, via TLDR22 SEP 2026as publishedR
  6. 06$1 billion...development slowdown, with each side committing at least $1 billion over five yearsAnthropic, via TLDR22 SEP 2026as publishedE
  7. 07$100MICONIQ's Pacesetter Index shows AI-native startups hitting $100M ARR in quarters, but the sample is a hand-picked...ICONIQ, via TLDR22 SEP 2026as publishedR
  8. 087,764.70S&P 500 close, +1.49% on the sessionYahoo Finance22 SEP 2026yfinance closeR
  9. 0927,122.09Nasdaq close, +2.26% on the sessionYahoo Finance22 SEP 2026yfinance closeR
  10. 10$86,425.60Bitcoin close, +6.51% on the sessionYahoo Finance22 SEP 2026yfinance closeR
  11. 11$2,774.03Ethereum close, +4.96% on the sessionYahoo Finance22 SEP 2026yfinance closeR
  12. 12$92.29Crude Oil (WTI) close, -7.99% on the sessionYahoo Finance22 SEP 2026yfinance closeR
  13. 1378 / 100Crypto Fear & Greed close, Extreme Greed on the sessionalternative.me22 SEP 2026as publishedR
  14. 147.0 / 10Vapi, Koda score across four dimensions04R dossier22 SEP 202604R dossier scoringE
V
Verified: the stat gate corroborated this figure against an independent search before publication.
R
Reported: the figure is carried as published by the linked source and was not independently corroborated.
E
Estimated: the figure is approximate or hedged, either in the source or by the stat gate.

Method names the computation path behind the row, from a closed vocabulary: as published, yfinance close, 03B search corroboration, 04R dossier scoring, not recorded.

A figure is money, a percentage, a spelled magnitude, a multiple, a rate carrying its unit, a score over its denominator, a thousands separated number or an index close. Bare integers and years are not counted, so a phrase like "over 100 integrations" carries no mark. A chart's own scale line is the axis of the level above it and is counted once, with that level.

Get the morning Signal

179 editions so far, one a day. Unsubscribe anytime.

Forward this to one operator you work with. Your referral link is in every email; milestones at koda.community/refer.