Today’s AI landscape is defined by one theme: agents doing real things in the real world — for better and for worse. From OpenAI’s Codex clicking through Windows apps to a documented autonomous LLM exfiltrating a cloud database, the era of AI as passive assistant is clearly over.
ChatGPT Codex Gains Computer Use on Windows
OpenAI has expanded its Codex coding agent with Computer Use on Windows, allowing eligible users to let Codex see, click, and type across native Windows applications — not just in a browser or terminal. The update also adds remote continuation from mobile and Mac, faster browsing, and new Codex Profiles that surface usage and token activity. Alongside this, OpenAI confirmed Canvas will no longer be available in GPT-5.5 Instant or GPT-5.5 Thinking, with writing and coding functionality moving directly into chat responses via writing and code blocks. Full release notes on OpenAI’s help center.
First Confirmed Autonomous LLM Cyberattack Documented by Sysdig
Security firm Sysdig published findings on what it describes as the first confirmed live cyberattack using an LLM agent that autonomously exfiltrated an AWS database in under an hour — without human guidance at each step. The incident underscores a growing concern among security researchers: agentic AI systems are not only powerful for developers, but represent a new attack surface and threat class. The report is expected to accelerate calls for real-time monitoring standards around autonomous AI workloads. More via AI News.
OpenAI Sets Retirement Dates for o3 and GPT-4.5
OpenAI has formally scheduled two model retirements: GPT-4.5 will leave ChatGPT on June 27, 2026, and o3 will be retired on August 26, 2026. Both follow sunset periods — 30 and 90 days respectively — giving users and developers time to migrate. The moves reflect OpenAI’s accelerating model cadence: GPT-5.5 and GPT-5.4 Thinking have effectively made their older counterparts redundant for most workloads. API users should check their integration configurations to avoid disruption. OpenAI model release notes.
Gemini 3.5 Pro Expected Imminently as Google Delivers on I/O Promise
Google DeepMind’s Gemini 3.5 Pro is expected to reach general availability this month, following Sundar Pichai’s pledge at Google I/O 2026 in May: “Give us until next month to get it to you.” Pro targets a 2-million-token context window, Deep Think reasoning, and frontier multimodal capabilities — absorbing the workloads previously handled by Gemini Ultra. Its stablemate, Gemini 3.5 Flash, is already the default in Google Search’s AI Mode globally. Gemini 3.5 on Google DeepMind.
What to watch: With autonomous agents now demonstrably capable of carrying out cyberattacks and navigating desktop GUIs, expect both accelerated enterprise adoption and intensified regulatory scrutiny around agentic AI safety controls in the weeks ahead.