Extremely Severe Zero-Day Vulnerability in PeopleSoft — Hundreds of Organizations Likely Affected

A critical zero-day vulnerability has been discovered in PeopleSoft, the enterprise software owned by Oracle. According to reporting by Ars Technica, the vulnerability has been rated as close to 'the most severe imaginable,' and it is believed to affect hundreds of organizations. There are also reports of confirmed cases in which gigabytes of data were stolen.

PeopleSoft is an HR and financial management system widely used by large enterprises and government agencies, which is why the scale of the impact is a major concern. As of now, the specific details of the vulnerability and the status of a fix are unclear from the reporting, but Oracle is under pressure to respond quickly.

General-Purpose LLMs Outperform Medical-Specialized AI — Paper Published in Nature

A paper published in Nature Medicine has delivered an interesting finding to the AI industry: general-purpose large language models (LLMs) have outperformed AI models built specifically for the medical field on medical benchmarks.

This is a result that challenges the long-standing assumption that specialized domains require specialized AI. It shows that as the foundational capabilities of general-purpose models have improved, they can now surpass specialized models even in highly specialized fields such as medicine. That said, the paper is based on benchmark evaluations, and effectiveness in actual clinical settings will need to be verified going forward.

Ex-DOGE Staff Raise $130 Million for AI National Security Startup

According to Vanity Fair, former staffers of DOGE (Department of Government Efficiency) have launched a startup that applies AI to national security and raised $130 million (roughly 20 billion yen) in funding.

The application of AI technology to national defense and security is expanding rapidly in the United States, and people with in-government experience have been moving to the private sector to start companies one after another. This appears to be part of that broader trend, though ethical and legal debates are expected around former government officials moving into the private sector and the military use of AI technology.

Qwen and Gemma Beat Claude Mythos with Scaled test-time compute

A notable technical result was reported in Reddit's LocalLLaMA community: by scaling test-time compute, Qwen-3.6-27B and Gemma-4-31B outperformed Claude Mythos on code optimization and speedup tasks.

Scaling test-time compute is a technique that improves output quality by allocating more compute at inference time. The result is particularly interesting because it suggests that even relatively small open models, when they use this technique effectively, can deliver performance rivaling state-of-the-art closed models.

That said, this result comes from an individual verification by a community member, and the reproducibility and evaluation conditions still need to be checked in detail.

Benchmarking Claude Code's fast Mode — About 2x Faster with Quality Maintained, at 2x the Cost

A verification article published on Zenn put hard numbers on Claude Code's fast mode. Based on measurements over 120 trials, fast mode cuts processing time by a median of about 1.9x (and up to 2.34x under some conditions) while largely preserving quality, with the tradeoff that costs roughly double.

The estimate that you are effectively 'buying time at about $40 per hour saved' is a practical yardstick when deciding whether to use fast mode in day-to-day development. Note, however, that in 2 of the 120 trials fast mode switched back to the standard mode without notice, so that is something to keep in mind when using it.

Agentic AI Optimizes Real Estate Title Work on AWS

The AWS Machine Learning Blog featured a case study in which Rocket Close combined Strands Agents, Amazon Bedrock, Knowledge Bases, and MCP tools to optimize real estate title work (the process of verifying property ownership).

Title verification in real estate transactions has traditionally been a time- and labor-intensive process, but the company reportedly achieved operational efficiency gains and cost reductions by leveraging agentic AI. The post walks through the reasoning behind the technology choices, lessons learned, and the business impact in detail, making it a useful reference for anyone looking at practical agentic AI deployments.

---