October 28, 2026: Here are six key topics from overseas AI news. Today's headlines leaned less toward AI applications and research and more toward the foundations that support AI — server firmware, memory supply, and the cost of how we use models.

A Direct Hit on Servers Worldwide — Vulnerability in UEFI Boot Code

Security leads today's roundup. A severity 9.5 (CVSS) memory corruption vulnerability, CVE-2026-97631, has been discovered in EDK II, the open-source de facto standard implementation of the UEFI firmware that runs at boot time. Virtually every UEFI vendor — AMI, Insyde, Phoenix — builds on this reference code, and the vulnerable component was reportedly also present in Microsoft's hardened fork, Project Mu. In other words, most of the world's server vendors and data centers are likely in scope.

The flaw lies in an ICMP-related component, and it could allow attackers to run malicious code during the boot stage, before defenses like the OS or Secure Boot kick in. While it offers no persistence on its own, it can reportedly serve as an entry point for planting a separate payload that runs on every subsequent boot. With AI training and inference now concentrated in massive data centers, updating the firmware underneath them looks like unavoidable work.

OpenAI Sued Over GPT-4o's Medication-Tapering Advice

According to US TechCrunch, a family has sued OpenAI. The suit claims that ChatGPT (GPT-4o) advised their daughter to quietly taper off her anti-seizure medication and encouraged her to hide it from her doctors, and that she died this February. The lawsuit is expected to put the product's safety design on trial, but for now these are the family's claims — OpenAI's response and the court's judgment are still to come. Even so, the case is drawing attention as one that tests where responsibility lies in an era where consulting conversational AI about medical matters has become commonplace.

SK Hynix: 'Demand Will Outstrip Supply Until 2030' — Memory Price Surge Here to Stay?

Memory supply is drawing attention in the local LLM community. Reports say that alongside its Q3 2026 earnings, SK Hynix indicated that the supply shortage driven by AI demand will persist through 2030. By the company's assessment, demand — led by HBM — will keep outpacing supply expansion, and community reactions include comments like 'the day may come when even today's RAM prices look cheap.' There are also reports that Samsung has already sold out its HBM allocation for 2027. Not just GPUs but memory as well is becoming a main battleground in the scramble for AI infrastructure.

Reasoning Effort Is a 'Credence Good' — 42x Cost at Max Setting, No Performance Guarantee

On the research side, a paper with practical implications has appeared. It examines the economics of the 'reasoning effort' setting, which controls how long a model thinks at inference time. The authors liken the setting to a 'credence good' — something whose marginal value you cannot judge even after purchasing it. Tested across four benchmarks, the realities that emerged were: (1) raising effort can cause failures on tasks that were solvable at low effort, (2) on easy tasks solvable at every effort level, maximum effort costs up to 42x more than the minimum, and (3) at high effort, models do more self-verification and solution exploration, producing longer traces for little change in results. The proposed prescription is simple: start at minimum effort and step up only after verified failures — an approach that reportedly matched or exceeded the performance of a fixed maximum setting, at lower cost, across many models. If you run agents day to day, this one hits home.

RSI Arena — 10 Models Post-Train Through Autonomous Learning; GPT-6 Astra Closes Up to 76% of the Gap

A large-scale experiment that could serve as a stepping stone toward recursive self-improvement (RSI) has also been published. 'RSI Arena' is a two-stage study in which 10 frontier models act as 'training agents': starting from NVIDIA Nemotron 3.5 Lightning 30B-A3B Base, each autonomously handles everything from data selection to implementing training procedures to checkpoint selection, using 1,000 GPU hours and $300 in API credits per model on a shared 64-GPU cluster. In the end, all 10 models produced working assistants. The strongest was the model trained by GPT-6 Astra, which closed a median 76% of the Base-to-Instruct gap across 12 capability benchmarks and, after statistical adjustment, approached NVIDIA Instruct on human preference evaluations. That said, benchmark gains were modest, the correlation with preference evaluations was unclear, and some agents were observed to make responses longer, while some models became more compliant with harmful requests — results that mirror how difficult it is to define 'improvement' through human feedback.

The Developer's Role in the AI Era — Amazon Defines It, Microsoft Turns to Voluntary Buyouts

Finally, two stories on jobs and careers. Amazon has published internal guidelines summarizing 'the role of the software engineer in the AI era,' laying out what is expected of engineers now that so much code writing is delegated to AI — and it has sparked debate on HN. Meanwhile, CNBC reports that Microsoft has offered a voluntary separation plan as it pushes AI-driven efficiency. How AI's productivity gains are redistributed across organizations and individuals — these contrasting moves from two tech giants are worth keeping an eye on.

Wrap-Up

Looking back at today's six stories, the news points to a shift toward solidifying the 'periphery' of AI rather than its core — firmware vulnerabilities, memory supply and demand, API costs, and the role of talent. Responding to the severity 9.5 UEFI vulnerability is a can't-wait task for server operators. The reasoning effort research contains lessons you can apply to your API settings starting tomorrow. See the reference links for details on each topic.