OpenAI Builds a Secure Sandbox for Codex on Windows
On May 13, OpenAI published an article detailing the sandbox technology behind running its AI coding agent, Codex, safely on Windows. By combining file access controls with restrictions on network communication, the sandbox minimizes the risk of Codex making unintended changes to a user's system. Windows has long been a primary platform for developers, yet it has been regarded as an environment where sandboxing is particularly difficult. This effort marks an important milestone in OpenAI's push to seriously expand Codex's reach from Mac/Linux users to Windows developers.
"Incognito Chat" Arrives on Meta AI, Leaving No Conversation History
Meta has started offering an "Incognito Chat" mode for WhatsApp and the Meta AI app. According to Mark Zuckerberg, conversations in this mode are processed in a protected server environment, and even Meta itself cannot access their contents. Chat history disappears as soon as the session ends. Zuckerberg claims that "Meta is the first AI lab to offer this level of private AI usage."
With privacy concerns around AI assistants mounting, it is notable that a major platform is now offering a "don't retain data" option. However, technical details about the protected server environment, and whether any third-party verification exists, have not been disclosed at this point.
Apple Prepares to Distribute AI Agent Apps on the App Store
According to MacRumors, Apple is moving forward with plans to enable the distribution of AI agent apps on the App Store. The App Store's review guidelines currently assume conventional apps, and how to position autonomously acting AI agents remains an open challenge. If AI agent apps are permitted, apps that perform tasks on the user's behalf—automatic email replies, schedule coordination, delegated web operations, and the like—would enter the iOS/macOS ecosystem in earnest.
The safety standards and transparency requirements Apple sets in this area could influence AI agent norms across the entire industry. A detailed timeline and technical requirements have not yet been announced.
AI Chatbots Surfacing Real Phone Numbers Becomes a Problem
MIT Technology Review reports that AI chatbots, including Google AI (Gemini), have repeatedly surfaced real people's phone numbers to third parties. One Reddit user reported being plagued for about a month by calls from strangers saying they were "looking for a lawyer" or "looking for a product designer." In March, an Israeli software developer was contacted on WhatsApp after Gemini output his phone number as an erroneous customer service contact. In April, a PhD student at the University of Washington got Gemini to output a colleague's personal mobile number.
Experts warn that this is a new pattern of privacy violation by generative AI, and the problem is made more serious by the fact that users currently have no easy way to prevent it.
Tencent Significantly Expands AI Investment as China's Chip Supply Improves
According to The Decoder, Chinese technology giant Tencent plans to significantly expand its investment in AI infrastructure in the second half of 2026. The move comes as domestic Chinese semiconductor makers accelerate production of AI chips and chip supply improves. The company has also announced strong earnings for the first quarter, and is reportedly in talks to invest in the AI company Deepseek.
With semiconductor export restrictions continuing amid U.S.-China tensions, Chinese companies are growing increasingly dependent on domestic chips, and Tencent's expanded investment can be seen as a move that symbolizes this trend.
Most U.S. Physicians Quietly Use AI Tools
According to NBC News, many physicians in the United States have quietly adopted AI tools in their daily practice. Drawing particular attention is an AI tool called OpenEvidence, designed to give physicians quick access to the latest medical literature and clinical evidence. Many use it on the job without going through any official adoption process, showing that an "invisible spread" of AI use in medicine is underway.
While AI adoption in clinical settings boosts efficiency, it is also fueling debate over accuracy verification and where responsibility lies.
---
**Also worth noting:** AWS and Cisco have announced a partnership aimed at securing AI agents, jointly developing solutions that scale to MCP (Model Context Protocol) and A2A agent communication. Meanwhile, LessWrong research on "predicting rare LLM failures with 1/30th the rollout count" has been drawing attention.