Linux Foundation and 20 Major Tech Companies Launch "Akrites" — Fixing OSS Vulnerabilities Before They're Exploited
On June 26, the Linux Foundation announced "Akrites," an initiative to fix vulnerabilities in critical open source software before they can be exploited by AI tools, in partnership with roughly 20 tech companies, AI labs, and banks.
As AI-driven attack techniques grow more sophisticated and faster, the risk of known vulnerabilities lurking in existing OSS infrastructure coming under attack is rising. Akrites aims to prevent AI-enabled cyberattacks before they happen by systematically discovering and fixing these vulnerabilities.
At the same time, some argue that AI won't fundamentally change the cybersecurity game — it will only make attacks faster. An essay from the R Street Institute analyzes that AI's central role is accelerating existing attack patterns rather than creating entirely new threat types, which reinforces the importance of preventive approaches like Akrites.
Vercel Releases "AI SDK 7"
Vercel has released version 7, a major update to its AI SDK library for building AI applications. It's also generating buzz on Hacker News.
The AI SDK acts as middleware connecting frontends to AI models, abstracting away features like streaming responses and tool calling. Details on version 7 are available on the official blog, and the library continues to evolve toward making it easier for developers to integrate AI capabilities into web apps.
AI Search Engine "Exa" Raises $250M Series C
Exa, which provides an AI-native search engine, has announced the close of a $250 million Series C funding round.
Unlike conventional keyword-based search, Exa is built on an architecture centered on semantic search. It is increasingly used as the retrieval frontend for RAG (Retrieval-Augmented Generation) pipelines in generative AI applications, and this funding is expected to accelerate the expansion of both search quality and index scale.
Competition in the AI search space is intensifying, further fueled by the rise of vector search databases, and the structure in which well-funded players lead the market is getting stronger.
California Launches an AI Job Impact Tracker
The California state government has launched an "AI Job Loss Tracker" to track job losses driven by AI adoption. Bloomberg reported that the move is part of a policy response to growing layoff concerns.
The initiative turns data on which occupations and industries are affected by AI-driven automation into something visible, and it is regarded as one of the earliest examples of full-scale, state-level monitoring of AI's employment impact in the United States. It stands as a symbolic case of AI policy shifting from regulatory debate to responses grounded in real data.
Most recently, Oracle was reported to be cutting 21,000 jobs on the grounds of AI adoption, showing that AI's impact on employment is moving from abstract concern to concrete numbers.
The Frontier of Local LLM Optimization: Qwen 3.6 27B on Blackwell and Speculative Decoding
In the local LLM community, optimization know-how for real-world deployment is being actively shared.
A company shared its experience running Qwen 3.6 27B on llama.cpp in an RTX PRO 6000 Blackwell (97GB VRAM) environment. While the model delivers performance approaching Claude Sonnet levels, it ran into stability issues with sessions being interrupted during agent use in VS Code + the Copilot extension, sparking community discussion. They also reportedly achieved a 15–20% speedup by leveraging the MTP (Multi-Token Prediction) variant.
An article that interactively explains how Speculative Decoding works also drew attention. It is a key technique for accelerating inference and one of the important technologies that makes running large models in local environments practical.
In addition, a practical guide was published summarizing which model sizes actually run on different RAM capacities (8GB / 16GB / 24GB / 48GB). Its conclusions — that Q4 quantization is the best starting point for most personal use cases, and that 24GB or more of RAM is the rough benchmark for running 34B models comfortably — offer practical reference for hardware selection.