AI Sec Watch

President Trump signed an executive order creating a voluntary framework requiring AI companies to share their frontier models (cutting-edge AI systems at the technological frontier) with the federal government before public release, aiming to improve security and protect critical infrastructure. The order balances innovation concerns with security risks by directing federal agencies to develop a system for assessing the advanced cyber capabilities of AI models before they are released.

Microsoft announced MAI-Thinking-1, a new in-house AI model designed for advanced reasoning tasks, at its Build 2026 conference. The company claims this medium-sized model performs as well as leading models on software engineering benchmarks and was trained from scratch on clean data without using techniques from other companies' models. This represents Microsoft's growing effort to develop its own AI models instead of relying solely on its partnership with OpenAI.

Google is adding a feature to its Phone app that detects when scammers use AI to impersonate calls from people in your contacts list, alerting you so you can hang up. The feature is part of Google's June Android update, which includes several other security and convenience improvements across Android devices.

Microsoft is launching Microsoft Scout, a new AI personal assistant built on OpenClaw (a foundation model technology) that integrates into Microsoft 365 apps like Outlook, OneDrive, and Teams. Unlike the existing Copilot assistant, Scout can see and do more, functioning as a comprehensive personal assistant that helps employees with tasks like organizing calendars, managing expenses, and drafting emails.

Microsoft announced Project Solara, a new operating system (OS, the core software that manages a device) built on Android and designed specifically for gadgets that run AI agents (software programs that can autonomously perform tasks). The company demonstrated two prototype devices: a desk gadget similar to Amazon Echo Show with facial recognition, and a wearable badge with a camera and fingerprint scanner, both intended to provide access to AI agents.

President Trump signed an executive order requiring tech companies to voluntarily share new AI models with the federal government for review up to 30 days before public release. The framework aims to help identify national security and cybersecurity risks before powerful AI systems reach the public, though companies are not legally required to participate.

OpenMed versions before 1.5.2 have a remote code execution vulnerability (RCE, where attackers can run commands on the affected system) in how it loads privacy-filter models. The vulnerability exists because the software uses overly broad pattern matching on user-supplied model names, allowing attackers to trick it into loading malicious code from external sources. An unauthenticated attacker can exploit this by providing a fake model repository containing harmful code that gets executed with the same permissions as the OpenMed service.

Anthropic is expanding access to Mythos, an AI model designed to find software vulnerabilities, to 150 additional organizations across 15+ countries as part of Project Glasswing. The expansion includes industries like power, water, healthcare, and communications, though new partners must meet security requirements first. Since the initial launch in April with 50 partners, Project Glasswing participants have discovered over 10,000 high or critical-level security flaws.

Amazon Bedrock AgentCore is a tool that lets Software as a Service (SaaS) providers serve multiple clients, called tenants, with different security needs using the same AI agent. Resource-based policies (rules that control who can access a resource directly) let you grant some tenants cross-account access from their own AWS accounts while restricting others to traffic that stays only within a private virtual network, all without sharing credentials or creating separate user accounts for each tenant.