AI News Top 10
September 23, 2026
OpenAI and Anthropic shipped new frontier models three weeks apart, while a critical AI gateway flaw and LLM-driven malware dominated security news.
models
Anthropic and OpenAI ship frontier models within 21 days
Anthropic released Claude Opus 5.5 twenty-one days after Fable 5.1 and Mythos 5.1. OpenAI shipped GPT-6 Sol and Luna nineteen days after GPT-6 Astra. On September 12 Anthropic CEO Dario Amodei wrote that the industry should slow the pace of capability work. Opus 5.5 now tops the Artificial Analysis Intelligence Index at 58 points.
models
OpenAI halves the API price of GPT-6 Sol and Luna
OpenAI cut the price of GPT-6 Sol and Luna by 50 percent against the GPT-5.6 promotional rates. GPT-6 Sol (max) costs 1.06 dollar per task, down from 1.99 dollar, and Luna (max) falls from 0.18 to 0.07 dollar. Artificial Analysis reports almost no gain in raw capability.
robotics
Toyota plans 400,000 factory robots from 2028
Toyota will spend 6.42 billion dollar a year from 2028 to equip factories with 400,000 robots. Of those, 150,000 go into its own plants and 250,000 into group companies that make components and materials. The wheeled ELEY robot has two hands, and workers teach it tasks by wearing jigs shaped to its fingers.
security
Critical Bifrost AI gateway flaw allows command execution
Bifrost, an open-source AI gateway that routes requests to more than twenty LLM providers, carries a flaw that lets an unauthenticated attacker run arbitrary commands with a single HTTP request. It is registered as CVE-2026-90898 with a CVSS score of 9.8. The gateway holds the credentials of every connected model provider.
security
CLOSEDQUORUM malware runs command-and-control through LLMs
Cisco Talos describes CLOSEDQUORUM, a Windows implant that queries up to four LLM providers, among them Gemini, DeepSeek, Qwen and Mistral, to pick post-compromise actions such as stealing credentials and crypto wallets. Once installed, the malware needs no further commands from a human operator. Talos calls it the first publicly documented Windows implant using this approach for command-and-control.
agents
Rabbit returns with OS3, an agentic OS without its own hardware
Rabbit, the company behind the R1 device, launched OS3, an agentic operating system that runs in the cloud and operates over Windows, Mac and Linux machines. Users no longer need Rabbit hardware. The software approach comes two years after the company tried to bypass mobile apps with dedicated AI hardware.
hardware
Qualcomm runs a 30B model on the phone
Qualcomm announced the Snapdragon 8 Elite Gen 6 and the Snapdragon 8 Elite Extreme Gen 6, both built around on-device AI. The top chip runs a mixture-of-experts model with 30 billion parameters locally and supports 8K video. Motorola showed its Signature 27 flagship with the new chip at the Snapdragon Summit.
regulation
Dutch watchdog says spy services lack AI oversight
The Dutch review committee for the intelligence services (CTIVD) reports that the AIVD and MIVD have too little oversight of their own AI use. Both services process data and summarise and translate documents with systems they build themselves, without a clear AI policy. The committee concludes the services are not sufficiently in control and that decentralised development leaves oversight to individual technical staff.
research
Lab-grown neurons reach AWS as an AI substrate
The Biological Computing Company is bringing AI tools built on biological neurons to Amazon Web Services. The company writes a thin software layer over neuron circuits cultured in a lab and promises faster and cheaper AI video processing than conventional hardware. The move to a public cloud puts a long fringe field to a commercial test.
models
Anthropic cuts Opus 5.5 prices and falls back on safety
Anthropic says Opus 5.5 costs 40 percent less than Opus 5 on typical workloads at standard settings and generates output more than 30 percent faster. Tokens cost 4 and 20 dollar per million, and cache reads 0.20 dollar per million. At safety thresholds the system falls back to Opus 4.8 for cybersecurity and Opus 5 for biology and LLM development.
Key Trends
- • Frontier releases keep accelerating: Anthropic shipped three models in three weeks.
- • Competition moves to cost per task, not capability.
- • AI gateways and agent frameworks are the new attack surface.
- • Watchdogs want visibility into AI use before they can govern it.
- • On-device chips and alternative substrates pull inference away from the cloud.
Wat vond je van dit bericht?
