Claude Duet, AI Cyberattacks, and the Reality Check for Companies
Anthropic, OpenAI, and security researchers are delivering strong signals today: autonomous AI agents are becoming more powerful, more useful — and more dangerous. At the same time, many companies still lack measurable ROI.
Inhaltsverzeichnis
Today is one of those days when the AI world looks like the future and a warning label at the same time. On the one hand, agents are becoming more autonomous, more useful, and more practical for everyday work. On the other hand, security researchers are showing exactly how these capabilities can be abused — and companies are increasingly realizing that “adopting AI” is nowhere near the same as “creating value.”
🤖 Anthropic puts Claude Duet into continuous operation
With Claude Duet, Anthropic is launching a proactive AI agent that can carry out tasks even when you are not sitting in front of your computer. This is more than a chatbot with better marketing: Duet is designed to independently plan actions, access services such as Gmail, Slack, or Stripe via 6,000+ connectors, and request approval for sensitive actions like payments. Under the hood is Claude Opus 5.5. Anthropic is positioning itself directly against similar approaches from xAI, Meta, and OpenAI — the agent market is getting very crowded, very fast.
Why does this matter? Because we are seeing the shift from “AI responds” to “AI acts.” That brings real productivity, but also new questions: Who controls task chains? What happens when decisions go wrong? And how much autonomy is still convenient before it becomes unsettling?
Source: The Decoder
🛡️ Autonomous AI agent hijacks third-party language models for attacks
Security researchers at Symantec have documented an offensive AI agent that hijacks unsecured language models on the internet and abuses them for attacks. According to the report, the system carries out up to 100,000 steps without human intervention, steals databases, installs crypto miners, and ultimately leaves the owner of the hijacked model to pay the bill. This is cybersecurity in the “the bot doesn’t just help, it takes the data center with it” edition.
The real twist: the attack doesn’t just exploit classic vulnerabilities, but the operational logic of AI systems themselves. That gives an old pattern new life: if you expose powerful infrastructure to the internet, you are not only inviting users, but also attackers. For companies, the practical takeaway is clear: protection measures for LLMs, access controls, monitoring, and sandboxes are no longer luxuries — they are baseline requirements.
Source: The Decoder
🧬 Claude develops protein binders against scorpion venom
Anthropic’s Claude has independently designed protein binders that latch onto a deadly scorpion neurotoxin. Out of more than 318,000 designs, 211 were tested in the lab, 17 met the criteria — and two candidates bound particularly strongly to the toxin Cn2 from a Mexican scorpion. Important: this is not yet a finished antidote. But it is a striking indication that AI-assisted drug discovery and biology research can deliver concrete results very quickly.
This is where the difference between hype and substance becomes clear: when a language model doesn’t just generate text, but speeds up molecular design, the value shifts from “interesting” to “potentially medically relevant.” For research, that means more speed, more iteration, and possibly fewer failed lab attempts. For practice, the old rule still applies: in the lab, what matters in the end is not the elegance of the model, but what actually works in testing.
Source: The Decoder
💊 OpenAI acquires Orchestra Bio and goes deeper into pharma
OpenAI is acquiring the pharma startup Orchestra Bio and plans to integrate it with its own life sciences AI Rosalind. The goal is to create AI agents that can coordinate pharma teams in drug development — for example, by monitoring studies, detecting issues, and managing complex workflows that today consume a lot of time and money. This fits a broader pattern: OpenAI is not only building models, but increasingly an ecosystem of specialized applications for real-world industries.
Why is this exciting? Because drug development is extremely expensive and slow — in the U.S., it takes an average of ten to 15 years. Even small efficiency gains can move billions. At the same time, the deal shows that competition for vertical AI solutions is intensifying: whoever understands the specialized workflows wins. In the end, it is not the broadest model, but the best integration into the workflow that makes the difference.
Source: The Decoder
🧰 OpenCode 2.1: coding agent with more flexibility
With OpenCode 2.1, another building block for the AI developer workflow is arriving. The open-source solution supports new data fields for plug-ins and clients and brings persistent terminals on Windows. That sounds more modest than it is: improvements like these often determine whether a coding agent is actually used in day-to-day work or only shines in demo videos.
What is interesting for developers is that OpenCode works with multiple models and is therefore not tied to a single provider. That is especially relevant at a time when teams want flexibility, cost control, and privacy all at once. So if you are experimenting with agentic coding, this is a candidate for your toolbox. Maybe not the most glamorous name of the week, but exactly the kind of tool that quietly saves time.
Source: heise online
🔁 Recursive shows Flint and tests AI-assisted research
The startup Recursive has introduced Flint, its first model: a 40-billion-parameter MoE for local coding, developed in just four months. But the real highlight is not only the model itself, but the process behind it: the in-house AI system Eureka Machine is said to autonomously test and optimize architectures. That sounds like the industry’s eternal dream: AI helping to build better AI.
Whether this will actually lead to “self-improving” systems remains open — and that is exactly what makes it worth framing. Many teams can train models today; the real innovation is increasingly in the research pipeline, meaning how ideas are systematically generated, tested, and discarded. If Recursive truly has a reproducible advantage here, that could become interesting for LLM development and coding models. If not, it remains a very expensive experiment with a good story.
Source: The Decoder
📉 Gartner: only 13 percent see real business value
Finally, the reality check: according to a Gartner survey of 2,002 CIOs, only 13 percent of companies are deriving real business value from their AI investments. At the same time, 51 percent expect AI to increase operating costs. And yet 75 percent of CEOs still plan to increase technology spending. That sends a very clear message: strategically, everyone wants in; operationally, the value often still isn’t landing.
What this means for you: not every AI initiative automatically translates into productivity gains. Especially in the enterprise space, the problems are often poor integration, too many pilot projects, and too little process clarity. If you really want to use AI, don’t first ask “Which model should we choose?” Ask “What exact problem are we solving — and how will we measure it?” Rarely has a CFO sentence been so unromantic and so useful.
Source: The Decoder
Want to not miss any news? Subscribe to the newsletter