ChatGPT and Claude Shared Chats Abused for Malware Distribution — A New Security Threat
Attacks have been confirmed in which threat actors abuse the chat-sharing features of ChatGPT and Claude to spread malware. According to The Decoder, attackers embed malicious content disguised as error messages or installation guides within shared chats to lure users in.
At the root of the problem is the fact that these shared links are hosted on trusted domains such as chatgpt.com and claude.ai. Conventional security tools see them as "content from a trusted domain," allowing the malware to slip past detection. It's a clever attack technique that turns the convenience of AI platforms against users, making caution when opening shared links more important than ever.
OpenAI Codex Autonomously Operates PCs on Windows 11 — From Bug Discovery to App Testing
OpenAI's Codex app now supports the "Computer Use" feature on Windows 11, enabling AI to autonomously operate programs to test apps and discover bugs.
Users can start tasks and check progress from the ChatGPT mobile app even while away from their PC. The fact that an AI agent can recognize the PC screen and autonomously perform mouse and keyboard operations makes this a major step toward "PC operation automation" that goes beyond coding assistance. That said, autonomous operation comes with the risk of unintended actions, so it will be interesting to see how safety measures are designed for real-world use.
Salesforce Claims AI Agents Shortened a 231-Day Migration Project to Just 13 Days
Salesforce claims that after migrating its entire development organization to Anthropic's Claude Code, it completed a migration project originally estimated at 231 days in just 13 days. Running with no token limits, its April 2026 productivity metrics reportedly showed a 79% increase in pull requests per developer and a 5% decrease in incident rate.
However, these figures come from Salesforce itself and have not been independently verified. The developer community remains deeply divided over whether adopting AI agents is "a true revolution" or "the greatest accumulation of technical debt in history." That said, there is no doubt that this serves as a valuable data point as a concrete migration case study.
A Security Plugin for Claude Code Arrives — Real-Time Vulnerability Detection
On May 27, 2026, Anthropic released an official security plugin for Claude Code called "security-guidance." As explained in detail on Qiita, the plugin provides a mechanism in which one AI independently reviews code written by another AI, detecting and fixing vulnerabilities in real time.
As AI coding becomes more widespread, the importance of mechanisms that guarantee the security quality of AI-generated code grows by the day. Considered alongside the Salesforce case covered in the previous section, this kind of security layer looks set to become a must-have investment for companies seriously adopting AI-agent-driven development.
Fulloch V2 — A 100% Local Voice Assistant That Runs on 16GB VRAM
A fully local voice assistant called "Fulloch V2," which runs on a GPU with 16GB of VRAM, has been announced on Reddit's r/LocalLLaMA. It uses Qwen3.5-9B as the main LLM, with Qwen3-1.7B handling ASR and TTS, and supports real-time response speed and voice interruption (barge-in).
New in V2, alongside Home Assistant integration, is integration with local Obsidian vaults. Users can read, write, and search Markdown notes via voice commands, and it also offers semantic search powered by bge embedding models. Because it runs entirely locally, there are no privacy concerns — it looks poised to become a practical option in the smart home and personal knowledge management space.
AI Coding Is Not "Simple Cost Reduction" — What's the Fundamental Change?
An interesting analysis has been published on Qiita, arguing that adopting AI coding does not simply reduce costs — rather, it changes the very structure of the development process itself.
In traditional development, programmers absorbed the "loose ends" of requirements and specifications on the ground. In a "prompt creation → AI coding" workflow, however, requirements and specifications must be clearly defined upfront. This view holds that there will no longer be anyone to absorb ambiguity, suggesting that the programmer's role may shift from "the person who writes code" to "the person who clarifies requirements."
---