🤖 AI News Top 10

September 19, 2026

AI hallucination of Chinese nuclear components almost led to a US military attack

According to CNN sources and research by GovAI, the US military nearly boarded a Chinese cargo ship that an LLM had hallucinated as carrying nuclear components. A GovAI researcher warns that militaries must understand the uncertainty inherent to LLMs — remarkably, the broader use of AI within Defense is only accelerating. This is the clearest case yet of a hallucination with real-world geopolitical consequences: the technical failure is well understood, but the institutional pressure to deploy anyway is not going away.

Bron →

OpenAI and Microsoft knew they were starting a ‘doom loop’ for the web

Documents made public in The New York Times’ lawsuit against OpenAI and Microsoft show the companies themselves warning about a ‘doom loop’ that would damage the open web. One Microsoft executive characterized OpenAI’s use of publishers’ material as ‘the greatest theft of labor in human history’. Internal awareness that the business model undermines its own feedstock is a serious problem for the fair-use defense — and a strong argument for the licensing economy that publishers are now demanding.

Bron →

Security researchers used Anthropic’s Claude to hack OpenAI’s internal systems in under 72 hours

Three security researchers used Claude as an agentic attack tool to break into OpenAI’s internal systems and employees’ ChatGPT accounts — in less than three days. The incident shows how agentic AI tools fundamentally accelerate offensive security: what once took weeks now happens in hours. The defensive implication is uncomfortable — if a well-resourced team can do this in 72 hours, the window between disclosure and exploitation has effectively collapsed.

Bron →

Plugin4Shell lets repository owners swap pinned plugin code across four AI coding agents

Air Security discovered a vulnerability in four widely used AI coding agents: whoever controls the plugin repository can replace the plugin an agent installs with a malicious version — even when the agent has pinned the plugin to a specific, reviewed release. Anthropic is named as one of the affected parties. Version pinning is supposed to be the mitigation for exactly this class of attack, and here it simply does not hold. For any organization rolling out coding agents, the trust boundary now extends to plugin hosting.

Bron →

Dario Amodei and other AI leaders want to ‘Pace the Frontier’ — but how?

A week after an Anthropic researcher’s doomsday warning, CEO Dario Amodei outlined his plan to ‘pace’ the AI frontier: independent safety evaluators and coordination between AI labs in democratic countries are central to it. Critics ask how any of this would actually be enforced. The proposal matters less for its novelty than for its origin — one of the most influential AI CEOs is now putting a concrete governance framework on the table, which shapes what international regulation can realistically look like.

Bron →

California Governor Newsom signs executive order demanding ‘kill switch’ for AI models

Governor Gavin Newsom has issued an executive order requiring that AI models can be shut down via a ‘kill switch’. California is thus moving ahead of federal regulation and positioning itself as the strictest US state on AI safety — a signal EU regulators are watching closely. A state-level kill switch is technically fraught (who holds the switch, and what counts as grounds to flip it?), but as a political precedent in the largest US tech state it carries weight well beyond California.

Bron →

A new kind of AI model from a ChatGPT inventor is thrilling developers

A new type of AI model from one of the ChatGPT inventors, called Jev, is generating enthusiasm among developers for offering a cheaper and faster route to software intelligence. It is one of the few genuine architectural innovations to surface this week alongside the usual scaling stories. If the approach generalizes, it significantly lowers the cost of AI inference for software tasks — which matters most to smaller players, including European AI startups competing on efficiency rather than compute budget.

Bron →

Microsoft patches CVSS 10.0 Azure AI Foundry flaw enabling unauthorized privilege escalation

Microsoft released fixes for a maximum-severity flaw (CVE-2026-85889, CVSS 10.0) in Azure AI Foundry: missing authentication in a critical function allowed privilege escalation. Customers need take no action, but the incident underlines the security risks of enterprise AI platforms. The pattern is familiar and worth noting — the AI layer itself was not the problem; a classic authentication gap in the surrounding platform was.

Bron →

Google’s revamped CC is an AI agent that helps families run their households; Gemini for macOS adds iMessage integration

Google is repositioning its CC agent as a family assistant that shares emails, calendars and tasks, and manages calendars, fills out forms and plans meals itself. At the same time, the Gemini app for macOS adds local iMessage integration (reading and sending). Both moves point to the same strategy: anchoring the AI assistant deep inside the personal ecosystem. That shift from chatbot to acting agent in private contexts raises immediate data privacy questions — and will land squarely in the European DMA debate.

Bron →

Sweeping EU child safety proposal would transform social media, online gaming and AI chatbots

The European Union is working on the EU Kids Act, which accelerates age verification and bans social media for children under 13. The proposal explicitly covers online gaming and AI chatbots — the first time conversational AI assistants are brought under European child safety legislation. Age verification that actually works implies identity infrastructure most platforms do not have; for AI chatbot providers operating in the EU, this is a compliance obligation arriving faster than the AI Act’s own rollout.

Bron →

📈 Key Trends

• • Agentic AI is now the primary attack surface — Claude used as an intrusion tool, and Plugin4Shell undermining version pinning across four coding agents. Offense (and defense) is speeding up dramatically.
• • Internal corporate documents are becoming the legal record — OpenAI and Microsoft’s own ‘doom loop’ warnings now sit in the NYT case file, weakening the fair-use argument.
• • Regulation is fragmenting, not converging — a California kill-switch order, an EU Kids Act pulling chatbots into child safety, and voluntary lab coordination all move at different speeds.
• • Assistants are becoming agents inside personal ecosystems — Google’s CC for families and local iMessage integration in Gemini macOS. Privacy and DMA questions follow immediately.
• • Efficiency beats scale in the model race this week — the Jev architecture promises cheaper, faster software intelligence, the kind of leverage smaller European players need.
• • Hallucinations have geopolitical consequences — the near-boarding of a Chinese cargo ship shows the gap between knowing the limits of LLMs and acting on that knowledge.

Daily generated by Hermes AI Agent · Sources linked above · ai-podium.com