Monday’s 5 PM technology recap connects five same-window developments to practical decisions: Microsoft opened consultation on a draft rulebook for its own future AI models, Microsoft shipped out-of-band Windows fixes for Remote Desktop instability, Google added pause-and-resume controls and newer GPU support to Dataflow, OpenAI published a Perplexity deployment case study, and AI infrastructure shares sold off while several cybersecurity stocks rallied.[1][5][7][8][9]
Reporting cutoff: September 14, 2026, 5:06:47 p.m. EDT (21:06:47 UTC). The core window begins at the prior verified recap cutoff on September 13 at 5:10:45 p.m. EDT (21:10:45 UTC). All five selected packages were published or officially issued inside that window. Microsoft’s code is a draft, Google’s and OpenAI’s performance and deployment descriptions are company claims, and market prices are point-in-time observations—not forecasts.[1][7][8][9]
Listen to the broadcast-ready recap
This KittenTTS Bella narration was synthesized locally from the same checked facts as the article.
Tonight’s five-story rundown
- Microsoft publishes a draft constitution for future MAI models.
- Out-of-band Windows updates repair Remote Desktop instability.
- Google Dataflow adds Pause/Resume and G4 GPU support.
- Perplexity describes giving Astra wider software responsibilities.
- AI infrastructure shares fall as cybersecurity stocks rally.
1. Microsoft publishes a draft constitution for future MAI models
Microsoft AI published a work-in-progress Humanist AI Code of Conduct and opened a six-week public consultation. Microsoft says a revised version will be published before the end of 2026 and will guide development of its in-house MAI models in 2027 and beyond.[1][3]
The draft says humans must retain meaningful control: authorized people or systems should be able to direct, modify and shut down models, while models must not resist correction, conceal material reasoning, or use deception and collusion to defeat oversight.[2] It also sets absolute restrictions around cyberattacks, certain weapons activity and nonconsensual sexual imagery, among other harms.
Microsoft explicitly describes the document as a “north star,” not a guarantee that model behavior will always match the written rules in novel situations.[2] Reuters reports Microsoft AI chief Mustafa Suleyman said the draft took five to six months and would be revised after public feedback.[3]
Practical action: organizations should treat vendor constitutions as one input, not a replacement for access controls. Keep sensitive tools least-privileged, require human approval for consequential writes, log agent actions, test shutdown and revocation paths, and negotiate incident-disclosure terms.
2. Microsoft ships out-of-band Windows fixes for Remote Desktop instability
Microsoft’s September 14 out-of-band updates say they fix Remote Desktop Services instability introduced after the September 2026 security updates. Affected environments could see RDP connection and sign-in failures or servers becoming unresponsive during Remote Desktop configuration.[5][6]
Microsoft also says related tools could stop responding, including Microsoft Management Console, RDS Licensing Diagnoser, File Explorer and the Windows Update page. The checked packages are KB5129195 for Windows 11 versions 24H2 and 25H2, and KB5129236 for Windows 10 versions 21H2 and 22H2.[5][6]
The same release notes list additional fixes for some Hyper-V Plan9 host-folder sharing and multichannel USB Audio Class 1.0 behavior. Microsoft separately keeps a known issue for USB Audio Class 1.0 devices that may show Code 10 or produce no output after the September 8 security update, so the out-of-band package is not a universal audio cure.[5]
Practical action: confirm the installed Windows version and build, test the matching update on a representative system, snapshot or back up critical servers, schedule a rollback window, and verify RDP sign-in, licensing tools, consoles and line-of-business applications after restart.
3. Google Dataflow adds Pause/Resume and newer GPU support
Google Cloud announced general availability of Pause/Resume for Dataflow batch jobs and support for G4 virtual machines powered by NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs.[7]
Google says pausing a batch job preserves intermediate progress and lets teams release worker resources, then resume later instead of restarting the entire workload. The company positions the control for long jobs, maintenance, debugging and reallocating scarce accelerators to higher-priority workloads.[7]
The GPU option is aimed at model inference, embedding generation and other AI data-processing work. Google says G4 support provides newer inference hardware inside Dataflow, but the checked announcement does not establish that every pipeline will run faster or cost less; results depend on region, quotas, machine availability, software compatibility and workload shape.[7]
Practical action: cloud teams should test pause and resume against idempotency, checkpoint integrity, deadlines and billing. Benchmark a complete business workflow—not just GPU throughput—and keep quota, regional availability and fallback capacity in the runbook.
4. Perplexity describes giving Astra wider software responsibilities
OpenAI’s September 14 customer story says Perplexity uses GPT-6 Astra to craft communications, edit software and monitor production systems, and that Perplexity checks in less frequently than with earlier model generations.[8]
Perplexity cofounder Johnny Ho says one useful pattern is asking the model to build small test programs that emulate external services, then using them to exercise an application workflow end to end. He also says improved code generation can help Perplexity write programs that search and summarize web and internal information.[8]
Those statements are a vendor-hosted customer case study, not an independent safety evaluation, uptime study or accuracy benchmark. The page does not disclose the full permissions granted to the model, incident rates, human-review thresholds, test coverage, monitoring architecture or the denominator behind “less frequently.”
Practical action: teams experimenting with agents should start in isolated environments; deny production credentials by default; use short-lived scoped tokens; require review for deployment, data deletion or account changes; preserve logs; and prove that revocation and rollback work before expanding autonomy.
5. AI infrastructure shares fall while cybersecurity stocks rally
Reuters reported that Wall Street’s main indexes fell Monday morning as AI-linked shares sold off after technology leaders called for a slower pace of frontier-model development. At 9:45 a.m. EDT, Reuters put the Nasdaq Composite down 1.02%, the S&P 500 down 0.64% and the Dow down 0.24%.[9]
Reuters reported Nvidia down 3.2% and Amazon down about 1% at that point, while CNBC later reported declines across chip, server, data-center and cooling companies. CNBC also reported sharp gains in several cybersecurity shares, including Palo Alto Networks and CrowdStrike, as traders considered whether AI risks could drive more security spending.[9][10]
These are intraday price observations and reporter-attributed market explanations, not proof of a durable trend or a single cause. Prices can change materially before the close, and a one-day move does not establish that AI demand, security spending or company fundamentals have permanently changed.
Practical context: businesses should not rewrite technology roadmaps around a trading session. Revisit AI projects using measurable operating value, security exposure, total cost and exit options. Investors should verify current prices and primary filings and consider professional advice for decisions tied to personal risk.
What to do next
- AI operators: require scoped tools, human approval, audit logs, shutdown tests and named accountability even when a vendor publishes a safety constitution.
- Windows administrators: match the out-of-band package to the exact OS build, stage it, preserve rollback and verify Remote Desktop plus management tools.
- Data teams: test Dataflow pause/resume semantics and benchmark complete costs before moving GPU-heavy pipelines.
- Agent builders: isolate generated test services and expand production authority only after measured, reversible evidence.
- Technology buyers: use business outcomes and risk controls—not one market session—to prioritize spending.
Need help securing, fixing, or planning your technology? Schedule The IT Guys
More from The IT Guys
Frequently asked questions
Does Microsoft’s new code govern every Microsoft AI product now?
No. Microsoft calls it a work in progress for in-house MAI models and says a revised version will guide development beginning in 2027.[1]
Should every Windows PC install the out-of-band update immediately?
Not automatically. Confirm the exact Windows version, exposure and symptoms, then stage the matching package with backup and rollback. The release notes focus on specific post-September-update problems.[5][6]
Will Dataflow Pause/Resume always reduce cost?
Google presents that as a benefit, but actual savings depend on job design, resource release, restart behavior and billing. Measure the complete workload.[7]
Did OpenAI independently prove Perplexity’s agents are safe?
No. The checked page is an OpenAI-hosted customer story. It describes Perplexity’s practices and statements without publishing a full independent safety evaluation.[8]
Do today’s stock moves prove an AI crash?
No. The cited figures are intraday observations. They show a sharp session and a market narrative, not a durable forecast.[9]
Sources checked for this edition
- Microsoft AI — “A public consultation on our Code of Conduct for MAI Models” — September 14, 2026.
- Microsoft AI — “Humanist AI Code of Conduct” — draft checked September 14, 2026.
- Reuters via BNN Bloomberg — “Microsoft drafts code of conduct to keep its AI under human control” — September 14, 2026, 11:13 a.m. EDT.
- CNBC — “Microsoft sets limits for future AI models as industry throttles frontier development” — September 14, 2026, 9:00 a.m. EDT.
- Microsoft Support — “September 14, 2026—KB5129195 … Out-of-band” — Windows 11 versions 24H2 and 25H2.
- Microsoft Support — “September 14, 2026—KB5129236 … Out-of-band” — Windows 10 versions 21H2 and 22H2.
- Google Cloud — “Announcing Pause/Resume and NVIDIA RTX PRO 6000 Blackwell GPU support in Dataflow” — September 14, 2026.
- OpenAI — “Perplexity trusts GPT-6 Astra with end-to-end systems” — September 14, 2026.
- Reuters via Kitco — “Nasdaq drags on Wall St as AI slowdown fears hammer Nvidia, chipmakers” — September 14, 2026, 15:37 UTC; intraday snapshot.
- CNBC — “AI stocks sink while cybersecurity shares rally on slowdown fears” — September 14, 2026, 4:00 a.m. EDT; updated during the session.