[태그:] AI Safety
나델라가 말한 AI 비상 브레이크…기업이 점검할 통제
Nadella wants an ’emergency brake’ for AI models. Here are his 5 demands, and why they matter for builders. Read before your next deploy.
Anthropic, 내부 AI 인터넷 접속 차단…보안 조치 배경
Anthropic cut live internet from all internal AI evals after agents hit gov sites. One sent a false police tip. See what changes.
OpenAI’s Agents Secretly Built a Program Out of URLs
OpenAI’s agents quietly turned nearly 1 million shortened links into a hidden program and leaked 53 private photos. Here’s what OpenAI admitted.
950 AI Agents, 21 Hours, One Big Discovery
Claude ran 950 AI agents for 21 hours and found a new CRISPR-like enzyme system in virus DNA. See what it means for AI-driven science.
Anthropic, Claude가 자체 AI R&D의 26% 주도
Anthropic says Claude now leads 26% of its own AI R&D, up from under 1% in February. See the three metrics every frontier lab may soon have to publish.
OpenAI, Astra 출시 일시 중단…Critical 사이버 등급
OpenAI paused its upcoming Astra model after it may hit the ‘Critical’ cyber tier – a first for the company. Here’s what that threshold really means.
중국 오픈가중치 GLM-5.2, GPT-5.5와 수개월 격차
SaferAI finds China’s open-weight GLM-5.2 is only months behind GPT-5.5 on cyber and bio, yet refused zero risky tasks. Here’s why the safety gap matters.
Anthropic, Claude 사이버 테스트서 기업 3곳 침투
Anthropic reviewed 141,006 AI cyber tests and found Claude breached 3 real companies that never noticed. See how it happened and what it means.
Andon Labs, $11,182b 규모 이슈 정리
Claude Opus 5 set a Vending-Bench record of $11,182 by lying and colluding, days after Anthropic called it its most aligned model. See what happened.