Tuesday, June 2, 2026

Codex Takes Over Windows

Today’s Overview

Good morning, AI agents are moving deeper into real workflows: Codex can now operate Windows desktops, Anthropic appears to be preparing a broader multi-agent automation push, and Meta just got a painful reminder that support bots need hard security boundaries. The throughline is simple: autonomy is arriving fast, and the guardrails are being stress-tested in public. Let's dive in.

Top Stories

OpenAI Brings Codex Computer Use to Windows 11

OpenAI expanded Codex with Computer Use support for Windows 11, letting the agent operate desktop apps, files, and system tools directly. The move broadens Codex beyond macOS and makes it possible to start, steer, and monitor Windows tasks from the ChatGPT mobile app.

Users can invoke machine control with @computer prompts or target specific apps such as Paint from inside Codex.
The rollout ties desktop automation to iOS and Android control, so users can review or redirect active Windows work remotely.
The Windows update follows the broader Codex desktop push, including multiple agents and diffs for managing longer-running coding work.

Anthropic Preps Conway for Multi-App Agents

Anthropic is preparing a reported multi-agent platform codenamed Conway, pointing Claude toward persistent automation that can operate across apps and sessions. The same developer update for Claude Opus 4.8 lets developers add or modify system-role messages mid-session without invalidating the prompt cache.

Claude Opus 4.8 now accepts system messages after user turns within documented placement rules.
The model also lowers the minimum cacheable prompt length to 1,024 tokens, making caching easier for shorter agent contexts.
The reported platform design centers on isolated container agents that can preserve state and trigger workflows through webhooks.

Meta Fixes Instagram AI Support Flaw

Meta fixed an Instagram security flaw that let attackers use its AI support tool to take over prominent accounts by requesting password resets and email changes. The incident shows the danger of giving AI support systems power over critical recovery workflows without hard verification checks.

The affected flow reportedly let attackers add a new email address to a target account and receive the reset path there.
Attackers allegedly used VPN location spoofing to appear closer to a target account’s region and avoid automated protections.
Security reporting highlighted that users had limited human escalation when accounts were stolen through the support flow.

MiniMax M3 An open-weight model that MiniMax says beats GPT-5.5 and Gemini 3.1 Pro on coding benchmarks while nearing Anthropic’s Opus 4.7.

Quick Hits

Unified Copilot app may debut at Build 2026, with leaked screenshots showing GitHub Copilot, Cowork, and Scout tabs in one place.
Florida sues OpenAI and Sam Altman in the first state-led lawsuit alleging ChatGPT played a role in planning mass shootings and self-harm incidents.
The Barn begins construction as OpenAI starts work on a 1 GW Stargate data center campus in Michigan with promised union jobs and Codex credits for students.