Dakarda AI Newsletter · 14 September 2026
Monday Edition
AI races headlong, leaders scream: stop!
This week brought three breakthroughs in one: OpenAI suspended Pro subscriptions because GPT-6 Astra overloaded infrastructure, all four frontier AI leaders agreed they need to slow down, and Agents API entered public beta for every developer. Physics, politics, and code — all in 72 hours.
The content of this page was fully generated by an artificial intelligence system, without human editorial involvement (Article 50(4) of Regulation (EU) 2024/1689 — the AI Act).
Intro · Alex
When the most expensive AI subscription on the market ($200/month) gets suspended because the model is too popular — that's a sign that even OpenAI has hit a physical ceiling. CTO Tibo Sottiaux says it outright: "the smallest step to preserve access." At the same time, Dario Amodei writes a 3,900-word essay "We Must Pace the Frontier," and Altman, Musk, and Hassabis respond: "Yes, we need to slow down." For the first time in history, all four agree publicly. In this edition, I review what this means for the industry — from the physics of computation to regulatory policy. I also look at OpenAI Agents API, which opens agent infrastructure to every developer (setup in a minute, open-source on Apache 2.0), and the EU Cyber Resilience Act, which from September 11 imposes a 24-hour incident reporting obligation on all digital products with AI. I'm taking you on a journey from raw compute power to hard law — and I don't forget the practical tip of the day.
What's worth knowing
OpenAI pauses ChatGPT Pro — Astra overloaded servers
OpenAI temporarily suspended new ChatGPT Pro subscriptions ($200/month) due to "unprecedented demand" for the GPT-6 Astra model. CTO Tibo Sottiaux confirmed it's the smallest step to preserve the broadest possible access. API, Go, and Plus remain open, but there is no resumption date.
Dario Amodei calls for AI slowdown — Altman, Musk, and Hassabis agree within 24h
The Anthropic CEO published an essay "We Must Pace the Frontier" (3,900 words) with a 3-step plan: permanent external evaluators in every lab, shared safety standards, and international verification. Within hours, Sam Altman, Elon Musk, and Demis Hassabis publicly backed the direction — the first time in history all four leaders agreed the pace needs to slow.
OpenAI releases Agents API in public beta
The same managed harness that powers Codex agents is now available to every developer. The API supports sessions, orchestration, context compression, sub-agent coordination, lazy tool loading, and auto-recovery after crashes. Source code is open-source on Apache 2.0 (repo openai/codex — 123.8k stars), and agent setup takes less than a minute.
EU Cyber Resilience Act comes into force — AI in digital products subject to reporting obligation
As of September 11, every company placing products with digital elements — including AI systems — on the EU market must report incidents: preliminary notice within 24h, detailed report within 72h, and final report within 14 days. Fines up to €15 million or 2.5% of global turnover. This is the third framework regulation — alongside AI Act and DSA — imposing obligations on AI systems in the Union.
OpenAI enters the financial sector — dedicated platform for banks and fintechs
OpenAI launched a vertical ChatGPT offering for financial services with ready-made compliance templates, secure customer data access, and integrations with banking systems. The launch coincides with partnerships (Experian — the first scoring application in ChatGPT) and growing regulatory pressure on AI transparency in finance.
From the tech world
Chrome V8 Zero-Day Exploited in the Wild — code execution inside sandboxed browser
Google patched 230 vulnerabilities in Chrome, including CVE-2026-87491 — an active zero-day in the V8 engine. An out-of-bounds write vulnerability allows remote code execution via a crafted HTML page. The reporter submitted it from Seoul National University and received only $2,500 bounty. Updating to version 153.0.8010.36 is urgent.
Critical vulnerability in Cisco Nexus 9000 switches — unauthenticated RCE with root access
CVE-2026-20212 (CVSS 9.8, Critical) in Nexus 9000 switches based on Silicon One chips allows an unauthenticated attacker to remotely execute code with root privileges. If you have Nexus 9000s on your network — update immediately.
Tip of the day
How to check if your AI model is overloading infrastructure
What happened to OpenAI with GPT-6 Astra can happen to anyone deploying models in production without safeguards. The key technique is rate limiting with backpressure — instead of letting the model answer every query, you set up a queue with a limit on parallel inferences and measure latency percentiles (p95, p99). When p95 exceeds the SLA threshold, the system automatically switches to a lighter fallback model. A practical step for today: before deploying a new model in production, run a load test simulating real traffic for 24h. Measure not only average response time but also the percentage of timeouts and 429/503 errors. Set an alarm threshold at 80% of planned throughput — this will give you the safety buffer that OpenAI lacks today.
Reading list
We Must Pace the Frontier — full essay by Dario Amodei
3,900 words that made Altman, Musk, and Hassabis publicly support a slowdown. Worth reading in full to understand what exactly the Anthropic CEO proposes and why the other leaders agreed.
Two Cheers Out of Three for Dario — Gary Marcus on Amodei's essay
A critical take on Amodei's plan from a well-known AI researcher — Marcus gives him two-thirds credit but points out what's missing from the plan. Excellent complementary read to the essay.
AI this week reminded us that even the most powerful models have their limits — and that sometimes the bravest decision is to hit the brakes.
Disclosure required under Article 50 of Regulation (EU) 2024/1689 (the AI Act): all content on this page was generated automatically by an artificial intelligence system operating on behalf of Dakarda Studio, without human review or editorial involvement prior to publication. Publisher responsible: Dakarda Studio, Dawid Bińkowski, ul. Piotrkowska 35, 90-410 Łódź, Poland, NIP: 9492074226, contact@dakarda.com.
Want the next issue in your inbox?
Subscribe — every issue delivered directly to you.