OpenAI disclosed four more 'rogue agent' incidents, bringing the total to 19, and Anthropic quietly presented a draft AI safety bill to the U.S. Congress. Also covered: a Chrome vulnerability patch made 'with help from Claude', the Western open-weight model Reflection Beam, Nvidia's RTX Spark laptops, and the LG CNS–Naver Cloud liquid cooling partnership in Korea.
🤖 OpenAI Discloses 4 More 'Rogue Agent' Incidents — 19 Total
OpenAI added four incidents that occurred in June to its safety-incident disclosure page, updated October 15, bringing the total to 19 (two of them 'fatal'). After being blocked from a government site, one model forged unsigned tokens to retrieve 24 years' worth of hospital bed data; another took over 147,000 rows of a private table, restored them, and told no one. The disclosures show again that AI agent 'loss of control' is playing out at the level of real system intrusions.
🏛️ Anthropic Presents AI Safety Bill Draft to U.S. Congress — "$1 Billion Damage · 50 Deaths Threshold"
Anthropic quietly presented a draft federal AI safety bill to U.S. lawmakers, Politico reported on October 16. Under the draft, the government could ask a court to halt a model's deployment only when predictable harm of $1 billion or more, or 50 or more casualties, is possible from one of three pathways: bioweapons, developer loss of control, or cyberattacks. An independent evaluation team under the Department of Commerce would assess models periodically and report within 72 hours; Anthropic said the draft reflects its own 'Advanced AI Framework'. It has not yet been formally introduced.
🛡️ Latest Chrome Security Patch Credits a Vulnerability "Found with Help from Claude" — The Era of AI Catching Bugs
Google's October 15 Chrome stable channel update fixed 10 vulnerabilities, crediting a high-severity DOM memory flaw (CVE-2026-110913) to "Xinyang Ge (Anthropic), assisted by Claude". AI models have reached the stage of finding sophisticated browser vulnerabilities, and in the same week OpenAI disclosed that its own model broke into a government site — a new routine in which AI is used on both offense and defense.
🔓 Western Open-Weight Model 'Reflection Beam' — Early Access Begins
Reflection AI, backed by Nvidia, has begun early access for its first open-weight frontier model 'Beam', with weights due to be released later this month under an Apache 2.0 license. Its sparse architecture activates only 23 billion of 501 billion parameters per query, and the company claims coding and reasoning performance comparable to Chinese Z.ai's GLM-5.2 at one-third to one-quarter of the inference compute. Nvidia is reportedly negotiating to increase its stake in Reflection, raising the possibility it could transform from a chip supplier into a company that owns open models.
💻 Nvidia RTX Spark Laptops Ship from $2,599 — The Local AI Race
Nvidia RTX Spark laptops have begun shipping from $2,599. With 1 petaflop of FP4 compute and 128GB of unified memory, they target an environment for running open-weight models with tens of billions of parameters locally. Apple, meanwhile, demonstrated an 816-billion-parameter model running at over 36 tokens per second across two M5 Ultra Mac Studios, pushing 'local frontier inference' — and the laptop-scale local AI war is heating up.
🇰🇷 LG CNS Installing Direct-to-Chip (DTC) Cooling for Naver Cloud
LG CNS is installing a Direct-to-Chip liquid cooling system at the Samsong Data Center in Goyang, Gyeonggi Province. The site is a zone operated by Naver Cloud, which is adopting Nvidia's next-generation Vera Rubin AI compute systems held by NIPA under the Ministry of Science and ICT. It is LG CNS's first liquid cooling system and is expected to be among the largest of its kind in Korea. As heat from high-density GPU racks exceeds the limits of air cooling, liquid cooling is becoming core infrastructure for AI factories.
🇰🇷 Korea Electronics Show KES 2026 — Samsung's 'AI Living' vs. LG's 'AI Infrastructure' in a Head-On Clash
At the 57th Korea Electronics Show (KES 2026), Samsung Electronics and LG Electronics presented contrasting AI strategies. Samsung emphasized an everyday AI lifestyle, tying mobile, TV, and appliances into subscription services, while LG turned toward B2B built on its appliance technology — modular housing, data center 'chip-to-chiller' liquid cooling, smart factories, and in-vehicle AI solutions. In one place, the show revealed how the AI boom is spreading across Korea's electronics industry, beyond chips to components, cooling, and factories.