Google's Gemini Spark Gains Access to All Your Personal Data — and It Forgot the Boyfriend While Planning a Birthday Party
Google has unveiled a new AI agent, Gemini Spark. Gemini Spark is designed to access users' email, documents, and calendar to provide comprehensive support for everyday tasks.
When WIRED's Reece Rogers actually used Gemini Spark to plan a birthday party, the AI combed through emails and the calendar and drew up a detailed plan. However, it turned out the guest list left out the most important person: the boyfriend. The episode is a telling illustration of a current limitation of AI agents — they can access your data, but they can't fully grasp the context.
Comprehensive access to personal data brings convenience, but it also raises privacy concerns. How much transparency and control Google offers will likely determine how widely the agent catches on.
Gartner Predicts 40% of Enterprises Will Demote or Retire Their Autonomous AI Agents
In a May 26 announcement, research firm Gartner predicted that by 2027, 40% of enterprises will demote or decommission their autonomous AI agents. The cause: applying uniform governance to agents with widely different risk levels.
Gartner points out that in their rush to deploy AI agents, enterprises are neglecting to design tiered governance proportional to each agent's security, compliance, and ethical risk. Applying the same rules to every agent creates a "governance paradox," in which low-risk agents become over-constrained while high-risk agents receive inadequate oversight.
As enterprises push ahead with full-scale AI agent deployments, this warning is likely to hit home. Sound, risk-based governance design could well decide the success or failure of AI agent investments.
OpenAI in Talks to Add Citigroup and JPMorgan to Its IPO Bank Lineup
OpenAI is reportedly in talks to add Citigroup and JPMorgan Chase to the group of underwriting banks for its initial public offering (IPO), according to Bloomberg.
OpenAI appears to have already built relationships with major financial institutions, and the move suggests its IPO preparations are becoming concrete. A public listing by the biggest player in the AI industry could have a significant impact on the market as a whole.
That said, details such as the listing timeline and valuation remain unclear at this point. The sustainability of OpenAI's business model, the competitive landscape, and regulatory trends are likely to be key to how it is valued.
Spotify × Universal Music Officially Green-Light AI Covers and Remixes
Spotify and Universal Music Group have officially cleared AI-generated covers and remixes. When artists themselves grant permission, listeners will be able to rearrange their favorite tracks however they like.
The initiative also introduces a revenue-sharing scheme for artists: a portion of the revenue generated from AI covers and remixes will be distributed to the rights holders of the original tracks.
With concerns about the spread of "AI slop" music and copyright infringement looming over AI in the music industry, this counts as a positive step. This model, which combines rights-holder consent with proper revenue sharing, could influence other media industries as well.
Meanwhile, in an article titled "AI Slop Is Coming for Your Playlists," The Atlantic points out that a flood of low-quality generative AI music could threaten playlist quality, so quality control remains an open challenge.
Tiny-vLLM — A Fast LLM Inference Engine Built Entirely in C++/CUDA Goes Public
Tiny-vLLM, an open-source project that picked up 18 points on Hacker News, is drawing attention. It is a lightweight LLM inference engine implemented solely in C++ and CUDA with no Python dependency, and it touts fast inference performance.
The existing vLLM is Python-based and highly flexible, but that comes with overhead. Tiny-vLLM eliminates this in pursuit of higher throughput in production environments. Thanks to C++/CUDA optimization, memory management and batch scheduling are also more efficient.
The project is also valuable as an educational resource, offering a useful reference implementation for developers who want to understand the internals of LLM inference.
Cursor 3.6 Adds an "Auto-review" Mode — Enabling Safe Autonomous Operation for Long-Running Agents
Cursor 3.6, the latest version of the AI code editor, adds a new agent execution mode called Auto-review.
Previously, the Cursor agent required user approval every time it invoked tools such as Shell, MCP, or Fetch. In Auto-review mode, the agent decides and executes autonomously based on a set of rules, and the user reviews the results afterward.
This allows users to keep agents running on long tasks without stopping to approve every single action. That said, the design still requires confirmation for operations that carry security risks.
Dutch Authorities Take Down a Botnet Enlisting 17 Million Devices
Dutch police and the National Cyber Security Centre (NCSC) have jointly taken down a large-scale botnet that had roped in more than 17 million devices.
Acting on a tip from security researchers, they identified around 200 management servers hosted by a hosting provider in the Netherlands. The police seized several of the servers, and the provider took the botnet offline on the grounds that it was being used for criminal purposes.
The botnet is believed to have been linked to a Russia-based residential proxy network. Many cases involved IoT devices and routers being illicitly co-opted, and keeping devices up to date with security patches remains a critical ongoing challenge.
---