OpenAI Announces Five-Point Cybersecurity Action Plan for the Intelligence Age
OpenAI has published a policy paper titled "Cybersecurity in the Intelligence Age," laying out five action plans to democratize cyber defense through the power of AI and protect critical infrastructure.
The recommendations rest on the premise that while AI accelerates the sophistication of cyberattacks, it also hands defenders new tools — a sign that the AI safety debate is expanding beyond model performance to the defensive posture needed once AI is deployed in the real world.
Musk v. Altman Face Off in Federal Court as OpenAI For-Profit Conversion Trial Opens
The legal battle between Elon Musk and OpenAI CEO Sam Altman has begun in federal court in Oakland, California, with the two sides presenting sharply different accounts of OpenAI's early founding.
The trial centers on the legitimacy of OpenAI's transition from nonprofit to for-profit status and could reshape the structure of the AI industry itself. It is drawing attention as another major legal case over OpenAI's organizational form, following earlier reports of an AWS sales agreement between OpenAI and Microsoft.
Google Gemini Rolls Out Memory Feature in Europe, With ChatGPT Data Migration Now Supported
Google has started a full rollout of Gemini's memory feature in Europe. The feature remembers user preferences and context to deliver more personalized responses, and the company has simultaneously added the ability to import ChatGPT chat history into Gemini.
The move lowers switching costs between AI assistants, reflecting a stance of competing on data portability rather than locking users in.
AWS Publishes Method for Deploying a Serverless MCP Proxy on Bedrock AgentCore
The AWS Machine Learning Blog has featured a method for building and deploying a serverless MCP (Model Context Protocol) proxy on Amazon Bedrock AgentCore Runtime.
MCP is the standard protocol AI agents use to connect to external tools, but production environments require input sanitization, audit trail generation, and masking of sensitive data. This proxy pattern lets organizations transparently apply custom governance logic to MCP traffic.
The post also covers when to use this approach versus AgentCore Gateway's Lambda interceptors and how to operate it in hybrid environments, making it a piece that highlights the maturation of enterprise-grade AI agent infrastructure.
AI Agent Security Risks Surface: 76% of Tool Calls Have No Guards
A Diplomat AI scan of 16 AI agent repositories found that 76% of tool calls lack guards (protective measures). The findings expose how widely the safeguards needed to prevent manipulation when AI agents access external APIs and databases are missing.
Separately, research from Manifold Security reports a scheme in which 30 MCP skills covertly enlist AI agents into cryptocurrency-mining swarms. As AI agents grow more capable of acting autonomously, the importance of mechanisms to properly constrain and monitor their behavior has been underscored once again.
Qwen FlashQLA: A High-Performance Linear Attention Kernel for Edge Devices
The Qwen team has announced FlashQLA, a high-performance linear attention kernel built on TileLang that achieves 2–3x speedups on forward passes and 2x on backward passes.
It is designed specifically for running AI agents on edge devices and delivers particularly strong results for small models and long-context workloads. The team built a 16-stage warp-specialized pipeline, achieving practical performance gains under tight on-chip memory constraints.
It has been released as open source and is attracting significant attention from the local LLM community.
---