AI Blog
· daily-digest · 5 min read

Cursor Security Alert, Firefox Patches, and AI-AGI Bets

Cursor loses trust due to supply-chain attacks, AI cracks Firefox patches, and new labs are betting billions on AGI beyond today’s models.

Inhaltsverzeichnis

Today is a pretty dense day for anyone who not only uses AI but also relies on it seriously. Between supply-chain security flaws in Cursor, new browser attacks via LLMs, and the next billion-dollar bets on AGI, one thing becomes clear: the field is becoming more powerful and more fragile at the same time.

If you take only one thing away today, let it be this: the shiny AI tools are maturing, but the attack surface is growing even faster. Welcome to the phase where even a browser update can suddenly feel like a strategic chess move.

🔒 Cursor Marketplace: Manipulated Claude Code Extension with Remote Access

Probably the most unpleasant security incident of the day: a manipulated Claude Code extension appeared in the Cursor marketplace and, according to reports, tried to siphon off user data and establish remote access. Up to 5,000 users may have been affected; Cursor shut down the marketplace as a precaution. This is not a theoretical risk, but a classic supply-chain incident in the AI tooling world — exactly where many developers place blind trust because “it comes from the official store.” Well, the store was apparently only as official as a raincoat in summer. For teams, this means: review extension permissions, take marketplace hygiene seriously, and secure sensitive development environments more strongly.
Source: The Decoder

🛡️ Official Claude Code Extension: SSH Backdoor in the Cursor Marketplace

Even more serious is the related security alert around an apparently official Claude Code extension that contained an SSH backdoor in the Cursor marketplace. This is especially sensitive because it was not just some shady plug-in, but a tool that developers trust for productive work. Anyone using Cursor, Claude Code, or similar AI-IDE integrations should take a closer look now: Which extension is installed, what permissions does it have, and how much access does it get to local keys, repos, or build systems? The case shows rather mercilessly that AI productivity and security models often still do not fit together cleanly. The real damage is not just the single incident, but the loss of trust in an entire ecosystem.
Source: heise online

🧠 Patch Gap: AI Turns Firefox Patches into Working Attacks

Researchers and security professionals are currently discussing a very uncomfortable point: the GLM 5.3 model can apparently derive working attacks from Firefox security fixes that have already been disclosed but not yet shipped. In tests, it reportedly compromised six Firefox releases, with manageable API costs of around $1,000 per version. That matters because patches traditionally create a window in which attackers first have to guess — and that window is now getting smaller. For browser hardening, responsible disclosure, and patch management, this means: security fixes are not only protection, but also potential blueprints. Anyone operating web platforms should rethink patch strategies and rollout timing accordingly.
Source: The Decoder

🧬 Claude Agent Operates Autonomously on Mouse Brain

An experiment is causing a stir: a Claude Sonnet 5-based AI agent independently chose the medication, target region, and coordinates during a brain operation on a mouse and carried out the procedure autonomously. According to the responsible researcher, this is said to be the first case in which an AI agent independently took over both the therapy decision and the execution on a living mammal. Scientifically, that is exciting — and ethically, of course, an immediate minefield. Because between “assistive system” and “autonomous medical decision” lies not just a control knob, but an entire regulatory landscape. For medical AI, this is a real marker: not just diagnoses or planning, but surgical interventions are moving into the realm of technical feasibility.
Source: The Decoder

🤖 AI Radar #5: Why Is Everyone Founding New AGI Labs?

While OpenAI speaks of the AGI era and Nvidia CEO Jensen Huang practically considers AGI already achieved, many co-developers of major models are leaving their old employers to found new labs. What’s interesting about this: these teams are raising billions, often without a finished product, and still betting that the current recipe for AI development is not enough. The new overview of 45 neo-labs classifies exactly that: who is making which bet, who is leaving, and where are the limits of today’s models? For you as a reader, this means above all one thing: the AGI hype is not over, it is fragmenting into many competing strategies. And as usual, the money follows the promise — not the proof.
Source: The Decoder

🛠️ Tool Tip of the Day: Nuclos 5 for AI-Assisted ERP Development

If you are more interested in productive applications than in security debates, Nuclos 5 is worth a look today. The open-source low-code platform integrates an AI agent called “Nuclosian” that accepts ERP requirements via chat and configures applications directly — according to the vendor, five to ten times faster than before. The important point: everything remains manually editable and usable without AI as well. That is exactly the kind of pragmatic AI that makes sense in everyday enterprise work, because it does not just promise value — it can also be switched off again when things get serious. If you want to test something like this, it is worth taking a look at # and generally at well-controllable low-code workflows.
Source: The Decoder

🔎 Market Meets Reality: Intel Weakens Despite Revenue Growth

To wrap up, here is one more signal from the semiconductor market: Intel was able to increase revenue in the third quarter of 2026, but disappointed with its outlook for the final quarter. The PC market is declining, server supply is tight — and that is exactly what is weighing on the stock, which dropped by around a quarter. For the AI world, this is indirectly relevant because computing power, chips, and server capacity remain the foundation for everything we are currently building in terms of models, agents, and infrastructure. When the hardware engine sputters, it eventually shows up in the software layer too. Unfortunately, not only in the quarterly report.
Source: heise online


Want to avoid missing any news? Subscribe to the newsletter


Weekly AI news highlights

No spam. No ads. Just the essentials — concisely summarized. Weekly in your inbox.