Dakarda AI Newsletter · 16 September 2026
Wednesday Edition
Gemini 3.8, Agents and Market Rebellion — This Week in AI
Google launches Gemini 3.8 Live, Dreamforce showcases agentic CRMs, markets panic-react to a call for slowdown, and the EU AI Act begins to weigh for real — all in one week.
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
Welcome to a Wednesday that feels like a Friday, and we're only halfway through the week. Google has released Gemini 3.8 Live and Live Extended Thinking — voice models that outperform GPT-Live-1 Astra in independent tests and support 97 languages in real time. At the same time, Dreamforce 2026 kicks off in [ADDRESS] under the banner "Experience the Agentic Enterprise", with Salesforce betting everything on AI agents as a real business model. Meanwhile, Dario Amodei triggers the largest sector sell-off in months — the [ADDRESS] Semi Index drops 5.86%, and capital flees infrastructure for applications. In today's edition, we trace five events that paint a clear picture: AI is moving from the "everything, everywhere, all at once" phase into the "what works, what pays off, and what is safe" phase. Google wins the voice layer, Salesforce tests agentic monetization, the market reacts to real risks, and regulations start to bite financially. Plus in the tech section — the ClickFix attack looming in the air and the surprisingly simple hardware exploit DDRop breaking confidential computing. Sit back, it's going to be dense.
What's worth knowing
Google Gemini 3.8 Live — voice models outperform the competition
Google announced Gemini 3.8 Live and 3.8 Live Extended Thinking — audio models with low-latency WebSocket communication, support for 97 languages with seamless switching mid-conversation, and real-time video processing. Extended Thinking scored 82.6% on the AI Speech to Speech Index, outperforming GPT-Live-1 Astra and Grok Voice Think Fast 2.0. This is the third Gemini model in a month — Google is clearly accelerating and betting on voice as the primary interface.
Dreamforce 2026: Salesforce's agent revolution begins
The Dreamforce 2026 edition has kicked off in [ADDRESS] under the banner "Experience the Agentic Enterprise". The main theme is the Agentforce 360 platform — tools for building, deploying, and managing AI agents within CRM. Speakers include Marc Benioff, Andrew Ng, and the CEOs of OpenAI, Anthropic, and Nvidia. This is one of the biggest tests of whether generative AI actually delivers measurable productivity in enterprises.
AI and semiconductor sell-off after Dario Amodei's plea
The Philadelphia Semiconductor Index fell 5.86% after Anthropic CEO Dario Amodei called for slowing down the pace of AI development. Nvidia lost 3.6%, AMD 5.6%, Micron 6%, ARM 9.7%, ASML 7.2%, and SoftBank over 10%. Capital quickly moved to the application layer — ServiceNow gained 7.4%, Salesforce 4.7%, Coinbase 9.2%. The market is beginning to separate hype-driven infrastructure spending from real monetization.
EU AI Act: first real obligations and rising compliance costs
As of 2 August 2026, transparency requirements under Article 50 of the EU AI Act are in effect — companies must label AI-generated content in machine-readable formats and inform users of interaction with a model. The next deadline of 2 December 2026 will cover labeling of all synthetic audio, video, and text. Startup Reggy has entered the compliance market, and the Digital Omnibus amendment has pushed back deadlines for high-risk systems. For startups and developers, this is a signal that AI compliance is no longer optional, but a [ADDRESS].
Anthropic publishes report on state use of Claude
Anthropic published a Threat Intelligence Report, revealing that it blocked 16 accounts of an Iranian state actor who used Claude to create malware, phishing portals, and disinformation profiles. The report also describes a malicious profiling platform in Malaysia and an operator of a mass surveillance network for Mali's security services. This is further evidence of systematic, state-sponsored attempts to weaponize LLMs for influence operations.
From the tech world
ClickFix attacks infecting PCs and Macs are going viral
ClickFix (Fake CAPTCHA) attacks trick users into copying and executing a malicious command under the guise of CAPTCHA verification. Campaigns are intensifying, infecting both [ADDRESS]. This is one of the most effective [ADDRESS], exploiting user frustration and automation.
New DDRop Attack Breaks Intel TDX and AMD SEV-SNP
Researchers have developed a hardware attack called DDRop that uses a cheap ($159) DDR5 interposer to silently skip RAM writes in confidential computing environments. The attack breaks the fundamental isolation assumptions of Intel TDX and AMD SEV-SNP — there is no patch, because the problem lies in the hardware architecture.
Tip of the day
Verify the source before copying a command
ClickFix attacks show how easy it is to trick even a savvy user: just a fake CAPTCHA and a command to copy. The mechanism works on both systems and is hard to detect because it exploits trust in the website and clipboard automation. Practical step: never run commands copied from your browser unless you are absolutely certain what they do. Instead, check the CAPTCHA source — real verifications never ask you to paste code into a terminal. Use a password manager with built-in URL scanning and disable automatic script execution from the clipboard in your system settings.
Reading list
DDRop Attack: The Silent Killer of Confidential Computing
A detailed analysis of the DDRop attack that breaks the latest Intel and AMD security for $159. Required reading for anyone designing cloud architecture or thinking of confidential computing as a silver bullet.
Anthropic Threat Intelligence Report — September 2026
Anthropic's full report on state use of Claude — from Iran, through Malaysia, to Mali. It shows the scale and methodology of attacks on AI models that every AI developer should know.
The market is testing hypotheses, regulations are drawing boundaries, and [ADDRESS] are just beginning.
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.