Anthropic revealed the success of its defenses against model cloning ('distillation'), and Palantir and Ondas founded a sovereign AI company, 'Regna AI'. Also covered: the U.K.'s push for binding AI regulation, Meta's agent's high user retention, and Korea's AI memory boom.
🛡️ Anthropic Reveals Success Blocking AI Model 'Cloning Attacks' — Names Z.ai
Anthropic announced on October 12 that new defenses against rivals cloning (distilling) Claude models have cut extraction attempts by 90–99%. The result of restricting access to the reasoning process and blocking repeated queries routed through proxy networks — tens of thousands of related accounts have been suspended. Anthropic directly named Chinese AI startup Z.ai (Zhipu), alleging it tried to clone Claude while developing GLM-5.3, and warned of safety and national-security risks.
⚔️ Palantir × Ondas Found Sovereign AI Company 'Regna AI'
On October 12, Ondas and Palantir founded 'Regna AI', a sovereign AI company. It provides mission-specific fine-tuned AI models that customers own and operate within their own country or enterprise environment, working even on edge equipment — drones, stratospheric balloons, ground vehicles — where communications are limited. Ondas is the first customer, and the model-training infrastructure is based on Palantir's AIP platform. The White House also restored the U.K. AI Safety Institute's pre-deployment access to frontier models, a sign that international AI safety cooperation is picking up speed again.
🇬🇧 U.K. Pushes 'Binding AI Regulation' Targeting OpenAI, Anthropic, Google
The U.K. is pushing binding AI regulation that would impose mandatory security-incident reporting and independent model testing on frontier AI model developers. The move comes as the White House restored the U.K. AI Safety Institute's pre-deployment model access yesterday (October 12), a picture of international AI safety cooperation and regulation accelerating in tandem. The White House still backs voluntary safety standards, so the gap between the two approaches remains.
Meta's new AI agent 'Muse Wonder' reportedly posted 73% next-day user retention — roughly double OpenAI's ChatGPT. The agent is designed to handle tasks including legal document review, insurance claim processing, and software development. Accessibility issues have also been flagged, including problems with Amazon integration.
🇰🇷 Goldman Sachs Names SK Hynix, Samsung Electronics 'HBM Winners' — Market Up to $150 Billion by 2030
Goldman Sachs named SanDisk, SK Hynix, and Samsung Electronics the biggest beneficiaries of high-bandwidth memory (HBM) demand. The HBM market is projected to grow to $40–150 billion by 2030, and high-bandwidth flash (HBF) could cut AI inference costs by about 30%. Korean capital continues to concentrate its buying on Samsung and SK Hynix this year, while foreign investors sell heavily.
🇰🇷 Samsung Begins AI RAN Projects with KT and SK Telecom in October — Ministry's 'Hyper AI Network'
Samsung Electronics is starting AI RAN (radio access network) projects with KT and SK Telecom in October, part of the Ministry of Science and ICT's 'Hyper AI Network' program. Samsung is the sole global vendor for the KT project and the lead vendor for the SK Telecom project, building 5G standalone networks and performing AI-based network verification at industrial sites such as shipyards and petrochemical plants. The ministry views AI RAN, physical AI, and autonomous operations as foundational technologies for the transition to 6G.