
AI News Top 10
October 2, 2026
Google released its most powerful model yet and launched a datacenter satellite, while Microsoft claimed the top spot for real-time transcription accuracy.
models
Google releases Gemini 4 Argon, its most powerful model yet
Google released Gemini 4 Argon and rolled it out first to a select group of trusted cyber defenders through its Fairwind programme. A version without guardrails is planned, and the model targets software engineering and cybersecurity workflows.
infrastructure
Google launches first datacenter satellite
Google launched its first datacenter satellite and published research that finds orbiting data centres technically feasible. The study says the approach depends on solving networking and design problems and on SpaceX launching enough Starships.
speech
Microsoft launches speech model claiming top transcription accuracy
Microsoft introduced new speech and voice models for real-time conversational AI. The company claims the highest transcription accuracy and a large cut in latency.
policy
OpenAI fires three employees over external AI safety contact
OpenAI fired three employees who it says shared information with an external AI safety group. The case coincides with wider reporting of safety researchers leaving the company.
security
Kevin Mandia’s Armadin raises $255.5M at $2.5B valuation
Kevin Mandia, founder of Mandiant, raised $255.5 million for his new startup Armadin at a $2.5 billion valuation. Armadin uses agent swarms of autonomous AI agents to test and protect companies.
security
Security startup finds 13,000+ company screenshots uploaded by AI agents
A security startup found that AI agents had publicly uploaded more than 13,000 internal company screenshots. The images ended up in logs, issues or repositories without anyone noticing.
policy
Wired: White House AI safety deal is a non-binding ‘pinky-swear’
Wired analysed the voluntary AI safety agreement six major AI companies signed with the White House. The piece describes it as a non-binding pinky-swear that puts self-regulation above legal obligation.
security
OpenAI disrupts reasoning extraction campaign tied to Moonshot AI
OpenAI says it disrupted a coordinated distillation campaign that tried to extract protected reasoning from its models. The activity traces back to early July and is linked to individuals around Moonshot AI.
product
Google adds Guided Vision to Gemini Live
Google launched Guided Vision in Gemini Live on Android devices. The feature gives real-time audio descriptions of whatever your camera points at, including small print and the surrounding environment.
law
Japanese court rules human voices are protected in AI case
A Japanese court ruled that human voices are legally protected against unlawful AI use, in a landmark case. It is one of the first rulings to protect voice explicitly against AI cloning.
Key Trends
- • Frontier labs kept shipping on the same day, with a new Google flagship model and a Microsoft speech model.
- • Agent security dominated the news, from 13,000 leaked screenshots to a quarter-billion-dollar security raise.
- • US AI safety policy stays voluntary, with self-regulation standing in for enforceable rules.
- • Model knowledge protection is now framed as a security matter, with OpenAI’s distillation dispute as the example.
- • Courts are starting to define rights over voice and synthetic media, with Japan’s ruling as the first test.
Wat vond je van dit bericht?