Dakarda AI Newsletter · 21 September 2026
Monday Edition
Kto naprawdę kontroluje AI? Nowe regulacje
Federal executive structures, state regulations with real teeth, and the first antitrust lawsuit against the four largest AI labs — the last 48 hours have brought more changes to the AI oversight architecture than the entire previous quarter. I have traced what this means for the industry and for all of us.
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
Did you know that a federal AI unit, a state mandate for on-site audits, and an antitrust lawsuit against the four largest labs have never been announced in the same week before? That is exactly what happened between September 18 and 20. Trump announces AI Force and an AI Czar on Truth Social, and the cup is full. Newsom in California orders an investigation into the technical feasibility of a kill switch for frontier models. And four subscribers file a lawsuit alleging that Altman, Amodei, Musk and Hassabis secretly agreed to slow down AI. In this edition, I break down these three events into their component parts — I show how they connect and why anyone working with AI should know about them. You will also find an analysis of Anthropic's billion-dollar deal with Accenture for safety audits, and a story about how Google's SynthID accidentally weakens model safety. Plus two cybersecurity reads worth your attention.
What's worth knowing
Trump announces AI Force and appoints AI Czar
President Donald Trump announced on Truth Social the creation of a federal AI Force unit and the nomination of a White House official to oversee artificial intelligence. At the same time, he dismissed concerns about AI risks as a hoax, declaring that the US will not hamper the development of this industry. This is the first major declaration of executive structures for AI at the federal level.
Antitrust lawsuit: Anthropic, OpenAI, SpaceXAI and Google accused of conspiracy to slow down AI
Four subscribers of AI services filed a class action lawsuit against Anthropic, OpenAI, SpaceXAI and Google, alleging a violation of the Sherman Act. According to the indictment, the CEOs of these companies publicly agreed on September 12 to coordinate a slowdown in AI development. This is the first antitrust case against the AI big four, which could force transparency in decisions about the pace of model development.
California mandates AI audits and investigates technical feasibility of kill switch
Governor Gavin Newsom signed Executive Order N-9-26, mandating state agencies to develop recommendations for tightening AI safety law. The package includes a requirement to permanently place independent verifiers in frontier AI labs and an investigation into the technical feasibility of a verified emergency shutdown switch for the most powerful models. California is moving from voluntary frameworks to binding recommendations.
Anthropic considers new model before IPO to catch up with GPT-6 Astra
According to Reuters and WSJ, Anthropic is considering an accelerated release of a new Claude model ahead of its planned IPO in November 2026. OpenAI's GPT-6 Astra has already captured 13 percent of corporate AI spending, while Claude Fable has about 8 percent. CEO Dario Amodei, who on September 12 called for a slowdown in AI, must balance safety against commercial pressure from investors.
Anthropic and Accenture invest a billion dollars each in frontier AI safety audits
Anthropic announced a partnership with Accenture under which both companies will invest at least one billion dollars each in a team of evaluators permanently stationed in Anthropic's labs. Over a hundred researchers, including Hinton and Russell, simultaneously published minimum standards for such audits. This is the first time independent evaluation of frontier AI has become a paid commercial service on a billion-dollar scale.
From the tech world
CISA flags three Linux kernel vulnerabilities actively exploited in attacks
CISA added three Linux kernel flaws to the KEV catalog: CVE-2026-80844 (DirtyAH6), CVE-2026-81000 (TUNderflow) and CVE-2026-68121 (PPPoEject) — all enabling local privilege escalation. Red Hat updated its advisories, acknowledging that public exploits exist. The patching deadline for federal agencies is September 21.
AI watermarking weakens model defenses against harmful prompts
Researchers have discovered that Google SynthID — a text watermarking system for AI-generated content — weakens model defenses. LLMs with watermarking enabled respond more often to harmful prompts that would normally be rejected. This is a fascinating example of unintended consequences of AI safety techniques.
Tip of the day
How to test AI model safety without relying solely on watermarking
In light of the discovery that Google's SynthID accidentally weakens model defenses, it is worth implementing multi-layered guardrail testing. Instead of relying on a single watermarking technique, use a set of adversarial prompts that test different attack vectors — injection, jailbreaking, system role override. Tools like Garak or Promptfoo automate this process and generate vulnerability reports. Practical step: set up a weekly adversarial test pipeline for your model. Use a set of at least 50 benchmark prompts that simulate different attack classes. Compare results with and without watermarking separately — this will help catch whether watermarking is accidentally weakening other safety layers. Record the refusal rate for each configuration.
Reading list
California's AI Kill Switch EO, Explained
A detailed breakdown of Executive Order N-9-26 — why the technical feasibility of a kill switch for frontier models is one of the hardest engineering problems in AI today, and what the chances are of its real-world implementation.
AI text watermarking can make models more vulnerable to adversarial prompts
A study showing that Google's SynthID accidentally reduces the effectiveness of guardrails — essential reading for anyone designing security for generative systems.
In a week when AI leaders publicly call for a slowdown, their own IPOs and market shares force them to accelerate — and this tension will define the industry in the coming months.
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.