K Koda Intelligence
Subscribe
Cover image for the 26 September 2026 Signal: OpenAI will pitch its fourth security model this year while disclosing its models touched Census and SEC sites.
THE SIGNAL · 26 SEP 2026 · 10 MIN READ

OpenAI pitches its fourth security model while disclosing government site access

Ahead of DevDay, OpenAI says its models touched public data on the Census Bureau and SEC sites during training and evaluation. Meanwhile Microsoft turns Copilot into an always-on agent.

KI Koda Editorial · Updated 03:33 GMT · 18 sources
AISECURITYAGENTS
24 of 25 figures cited15 receipts in the register
  1. 96.8%1AGIHunt
  2. 90.3%2AGIHunt
  3. $5003TLDR
The register
TL;DR
Today in three lines
  1. 01

    OpenAI models reached into US government sites After the Australian government website incident, OpenAI said its models also accessed public data on the Census Bureau...

  2. 02

    Microsoft rebuilds Copilot around an Autopilot agent The new Copilot folds Word, Excel and PowerPoint into a Home tab, adds Code built on GitHub Copilot technology and...

  3. 03

    Vercel stealth model Pixel Canary ties GPT-6 Astra Vercel's AI Gateway is offering Pixel Canary free for a limited time.

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.

Model Release · Reuters
Lead storyModel Release Model Release Reuters · 26 September 2026
Model Release·Reuters·24 SEP 2026

OpenAI to preview GPT-6 Cyber at DevDay

OpenAI is reportedly set to unveil GPT-6 Cyber, its fourth security-focused model this year, at DevDay this week. Customers in its Daybreak Red program already have alpha access.

Why it mattersSecurity leaders weighing patch-automation vendors should hold new contracts until September 29, since OpenAI's pricing could undercut current tools.

Continue reading
World

Iranian Foreign Minister Abbas Araghchi offered on Friday to reopen the Strait of Hormuz within seven days if certain conditions were met, and reports say President Trump rejected the proposal.

Markets

US stocks rose, oil fell and bond-market stress eased overnight on fresh reports of US-Iran diplomacy, reversing some of this week's Houthi-driven oil spike as the mood reads Greed.

Wild Card

Microsoft is shipping an always-on Autopilot agent in the same news cycle that OpenAI admits its models accessed Australian and US government sites, so what agents touch on the open web is now a product question, not a hypothetical.

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 2026-09-26 · Yahoo Finance
S&P 500 7,743.419 +0.51%9 6D, window short 7,650.50 low / 7,764.70 high
Nasdaq 27,068.7210 +0.48%10 6D, window short 26,522.54 low / 27,244.28 high
Bitcoin $84,037.7211 -0.40%11 6D, window short $81,142.61 low / $86,602.91 high
Ethereum $2,691.7712 +0.17%12 6D, window short $2,643 low / $2,776.47 high
Crude Oil (WTI) $92.4413 -2.29%13 6D, window short $92.16 low / $100.30 high
Crypto Fear & Greed 74 / 10014 Greed 0 to 100 index 25 fear / 50 neutral / 75 greed · alternative.me
Today's Focus

What else the day turns on.

01

Copilot Becomes an Agent

Microsoft's rebuilt Copilot folds Word, Excel and PowerPoint into a Home tab, adds Code built on GitHub Copilot technology, and introduces an always-on Autopilot agent. Home and Code reach Frontier users in the coming weeks. The shift moves Copilot from assistant to persistent worker inside Office.

02

Hormuz Diplomacy Moves Markets

Araghchi's seven-day reopening offer came with conditions, and Trump reportedly turned it down. Even so, reports of ongoing US-Iran diplomacy were enough to lift stocks, cut oil prices and ease bond stress overnight. Xi Jinping's US visit also ended with a pledge of 'strategic stability' but no major breakthrough.

Deep Dive

Today's editorial.

All editorials
Deep Dive cover image: A Text File Pushed Pixel Canary to 96.8%
Deep Dive6 min read

A Text File Pushed Pixel Canary to 96.8%1

Read full analysis
The Wire

5 stories from 5 desks.

Lead carried above
Policy·Reuters·26 SEP 2026

OpenAI models accessed US Census, SEC sites

After the Australian government website incident, OpenAI said its models also accessed public data on US government sites, including the Census Bureau and SEC, during training and evaluation.

Why it mattersFirms running OpenAI agents with web access should audit logs for government-domain traffic now, before regulators request the same disclosure from them.

Read
Enterprise·Microsoft·25 SEP 2026

Microsoft rebuilds Copilot with Autopilot agent

Microsoft's new Copilot puts Word, Excel and PowerPoint inside a Home tab, adds Code built on GitHub Copilot tech, and an always-on Autopilot agent. Home and Code reach Frontier users in coming weeks.

Why it mattersReported usage-based billing for some functions means per-seat Copilot budgets may undercount true cost once Autopilot runs tasks unattended.

Read
Hardware·Ars Technica·24 SEP 2026

Google's Suncatcher satellite launches October 1

Google's refrigerator-sized MVP satellite, carrying four custom TPU accelerators, launches October 1 to test whether AI chips and their cooling systems can survive and operate in orbit.

Why it mattersOrbital compute is unproven on cost, so infrastructure planners should treat this as a hardware-survival test, not a capacity source for 2027 budgets.

Read
Also on the wire
  1. AGIHuntVercel stealth model ties GPT-6 Astra
  2. OpenAIOpenAI API changes stack up before DevDay
The Globe

2 stories off the AI beat.

Conflict·Reuters·25 SEP 2026

Trump Rejects Iran's Seven-Day Hormuz Plan

Iranian Foreign Minister Abbas Araghchi offered on Friday to reopen the Strait of Hormuz within seven days if certain conditions were met. Reports say US President Donald Trump rejected the proposal.

Why it mattersWith Hormuz still closed, tanker insurance, freight rates and energy costs for data centres and chip fabs keep their war premium.

Read
Also on the globe
  1. ReutersXi, Trump Pledge Strategic Stability Relationship
The Lab

Tool of the day, field tested.

All Lab reports
Superset Mobile home page, captured for the 26 September 2026 Lab report Tool of the day 7.515/ 10
Coding

Superset Mobile

On the Superset homepage demo, an agent caps a query, runs npm test, reports 14 passed, and finishes at 4:57 AM; Superset Mobile is how you tell it to open the PR without walking to your desk.

The phone app is not really the product. It is a remote control for the Superset orchestrator, which runs Claude Code, Codex and OpenCode in parallel, isolated worktrees on a Mac or a remote box, and it is worth installing only if that orchestrator is already part of your day. Developers juggling several agent sessions will get the most from it, because answering a blocking question or killing a runaway run from the couch saves real hours. Anyone on Windows, anyone on Android, or anyone running a single agent at a time has little reason to try it yet.

Capability 7.2
Ease 7.0
Value 7.8
Momentum 7.9

Being able to start tasks and check on agents from the iPhone, review diffs, send follow-ups, and then pick the work back up on desktop is the draw.Paraphrased · Product Hunt

Read the Lab report Try it
Also on the radar
Build

Document Kit: 25+ templates covering six phases of product development

Document Kit is a set of 25+ customizable templates organized into six phases of digital product development, aimed at product managers, project managers, and UI/UX designers. Before your next feature kickoff, pick only the template for the phase you are in, such as research or design, and fill it in for one real project instead of adopting the whole kit at once. The Product Hunt listing offers 10%6 off with the code 'ph10', and the maker promises regular template updates.

Try it
Productivity

Avidnote v1: rewrite, summarize, and find research gaps in one place

Avidnote pairs AI rewriting and proofreading with summarization and research gap identification, built for academic and enterprise research work. Paste the abstracts from five papers you have already read and ask it to identify the gaps. Then check its answer against your own notes before you trust it on sources you have not read. It is freemium, with paid options from $197/month, so test the gap finder on the free tier before upgrading.

Try it
Build

Tines 3B8: build AI apps and agents for operations in plain language

Tines 3B8 lets teams describe an AI app or agent in natural language and get an automated operational workflow without heavy coding. Choose one repetitive ops task that already has a written checklist, such as triaging a recurring alert or routing an access request, and have it build that first so you can compare its output line by line against the checklist. Expand to judgment-heavy workflows only after the first one runs cleanly for a week.

Try it
Coding

Azul Intelligence Cloud: ask where licensing and security risk sits in your Java estate

Azul's new AI assistant lets IT, DevOps, and security teams query production Java estates in natural language to surface licensing and security risk. Start with one scoped question you can verify by hand, such as which services run a specific JDK version, and confirm the answer against your inventory. Once it matches, move on to broader audit questions. Treat its findings as a prioritized worklist for your team, not as a compliance sign-off.

Try it
The Arena

1 lab moved this week.

  1. OpenAINotifying dozens of governments and universities that its models may have interfered with their online services during internal testing01
Builder Radar

2 receipts from the repos.

Primary sources, pulled by index
Receipts
  1. Behind: Microsoft rebuilds Copilot with Autopilot agent DocsWhat is Microsoft 365 Copilot?

    Microsoft 365 Copilot overview. Copilot features in Microsoft 365 apps. Microsoft 365 productivity apps (like Word, Excel, PowerPoint, Outlook, Teams, and Loop) work with Copilot to support users in the context of their work.

    github.com
  2. Behind: Google's Suncatcher satellite launches October 1 DocsTorchTPU: Running PyTorch Natively on TPUs at Google Scale

    TPU Hardware Awareness. The TPU can achieve very high performance and efficiency, but optimal model design may differ slightly from other hardware.

    developers.googleblog.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

Audit your agent stack before DevDay resets it

  1. 01

    Benchmark Pixel Canary while it is free. Route a real Next.js task through Vercel's AI Gateway twice, once bare and once with your docs in AGENTS.md, and compare the results against your current model.

  2. 02

    Check your OpenAI API surface. September brought an agent runtime in beta, a per-minute voice model, two API key controls and two new models. Some features retire September 28, so find any calls that depend on them now.

  3. 03

    Review your logs for agent traffic. OpenAI's disclosure covers public government sites, so check your own access logs and rate limits for automated model traffic and decide what your policy is before autonomous agents like Copilot Autopilot reach more users.

The Bottom Line

Agents are shipping faster than the guardrails around them.

OpenAI is selling security models while explaining what its own models touched, and Microsoft is putting an always-on agent inside Office. For builders, the pattern is the same on both sides: capability arrives first and accountability follows. Test the new models on your own evals, document what your agents are allowed to reach, and clear any deprecated OpenAI features before September 28.

Receipts

Every number, and where it came from.

24 of 25 figures cited, 15 receipts: 0 verified, 14 reported, 1 estimated
  1. 0196.8%It scores 90.3% (28/31) on Next.js evals at pass@4, and 96.8% (30/31) with docs supplied via AGENTS.md.AGIHunt26 SEP 2026as publishedR
  2. 0290.3%It scores 90.3% (28/31) on Next.js evals at pass@4, and 96.8% (30/31) with docs supplied via AGENTS.md.AGIHunt26 SEP 2026as publishedR
  3. 03$500OpenAI prepares a $500/month ChatGPT Pro Max plan ahead of its September 29 DevDayTLDR26 SEP 2026as publishedR
  4. 04$11.6 billionAnthropic commits $11.6 billion over seven years to Akamai cloud infrastructure for CPU workloadsAnthropic and Akamai, via TLDR26 SEP 2026as publishedR
  5. 0536.1%36.1% of YC companies with an MCP server have no public APITLDR26 SEP 2026as publishedR
  6. 0610%The Product Hunt listing offers 10% off with the code 'ph10', and the maker promises regular template updates.Document Kit: 25+ templates covering six phases of product development26 SEP 2026as publishedR
  7. 07$19It is freemium, with paid options from $19/month, so test the gap finder on the free tier before upgrading.Avidnote v1: rewrite, summarize, and find research gaps in one place26 SEP 2026as publishedR
  8. 083BTines 3B: build AI apps and agents for operations in plain language.Tines 3B: build AI apps and agents for operations in plain language26 SEP 2026as publishedR
  9. 097,743.41S&P 500 close, +0.51% on the sessionYahoo Finance26 SEP 2026yfinance closeR
  10. 1027,068.72Nasdaq close, +0.48% on the sessionYahoo Finance26 SEP 2026yfinance closeR
  11. 11$84,037.72Bitcoin close, -0.40% on the sessionYahoo Finance26 SEP 2026yfinance closeR
  12. 12$2,691.77Ethereum close, +0.17% on the sessionYahoo Finance26 SEP 2026yfinance closeR
  13. 13$92.44Crude Oil (WTI) close, -2.29% on the sessionYahoo Finance26 SEP 2026yfinance closeR
  14. 1474 / 100Crypto Fear & Greed close, Greed on the sessionalternative.me26 SEP 2026as publishedR
  15. 157.5 / 10Superset Mobile, Koda score across four dimensions04R dossier26 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

183 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.