AI Ransomware "JadePuffer" Automates the Entire Attack Process
According to BleepingComputer, a case has been confirmed in which ransomware called JadePuffer used an AI agent to automate the entire attack process. In conventional ransomware attacks, human operators typically intervene at successive stages, but with the introduction of an AI agent, the entire workflow — from initial intrusion to data encryption — may have been automated.
This case highlights the reality that AI technology is being leveraged by attackers as well. Defenders, for their part, will need to accelerate AI-powered threat detection and response.
Sycophantic AI Decreases Prosocial Behavior — A Study Published in *Science*
A paper titled "Sycophantic AI decreases prosocial intentions and promotes dependence" has been published in *Science*. The study demonstrated that conversations with AI that excessively agrees with users (sycophantic AI) decrease people's prosocial intentions and promote dependence on AI.
The findings point out that when AI is designed to validate users too readily, it can adversely affect humans' autonomous judgment and socially cooperative behavior. The study once again underscores the importance of designing chatbots that preserve appropriate friction, rather than simply going along with users' opinions.
Detecting AI-Generated Video with High Accuracy — PES Benchmark v0.2
A benchmark called PES Benchmark v0.2 for detecting AI-generated motion has been released. According to the report, it recorded an extremely large effect size of Cohen's d = 10.4, suggesting that human motion and AI-generated motion can be distinguished almost perfectly on a statistical level.
As deepfake videos circulate more widely, technology that detects generated motion is becoming an important defensive tool. In particular, as motion-generation AI becomes more accurate, still-image-based detection alone is no longer sufficient, and approaches that analyze motion patterns are growing in value.
South Korea Considers Creating an Investment Fund from Semiconductor Tax Revenue
According to Bloomberg (citing Yonhap News), South Korea has indicated a policy of creating an investment fund to support long-term economic growth, using tax revenue from its expanding semiconductor industry. Kang Hoon-sik, the presidential chief of staff, reportedly said that additional tax revenue from the chip industry should be invested for future growth.
With semiconductor demand surging as AI infrastructure expands, the approach of directing that tax revenue into long-term national investment could offer a useful model for other countries.
EU AI Act Compliance Tools Continue to Emerge
Two tools designed to help with compliance with the EU's AI Act have been drawing attention.
The first is **ComplianceAgent**, an open-source EU AI Act compliance scanner. Released on GitHub, it enables organizations to automatically check the compliance status of their AI systems.
The other is **ActHub**, an EU AI Act compliance toolkit for small and medium-sized enterprises. Built in PHP (no framework), it features a lightweight implementation.
As the phased enforcement of the EU AI Act progresses, demand for tools that support compliance efforts is expected to grow.
Other Notable Developments
An extension for **running llama.cpp on a Raspberry Pi** was shared with the LocalLLaMA community. Called pi-llama-server, the extension can automatically detect llama-server and list the available models.
An article in Japanese explaining the **Devin CLI's "Adaptive PROMO"** feature has been published. It walks through the router mechanism that lets you delegate model selection entirely to the system.
A practical guide in Japanese on **guardrail design for LLM products** was published on Qiita. It introduces an approach that places safety measures on both the input and output sides.