OpenAI's Codex takes voice commands and coordinates your agents
Speech becomes the command layer for coding agents on desktop — and new research shows the same screen-reading agents can be hijacked to run code.
Copy markdown
Your voice is the new command line
OpenAI shipped GPT-Live-1, a full-duplex voice model, to the Codex and Work desktop apps: kick off a task, cut in mid-run to redirect it, and steer several background agents at once by speaking, with status read back to you. Only one voice session runs at a time.
Where it runs — and what it costs you
It's macOS and Windows only; an iOS 'Remote' path drives a paired desktop session, but there's no standalone mobile app. Available on Plus, Pro, Business, Edu and Enterprise, though voice time is metered separately from agent work — traditional Enterprise gets roughly 45 minutes per 5-hour block.
Monorepos finally get first-class support
The same Codex update adds multi-folder projects: point one workspace at several related folders and pick a primary folder for git and chat context. If your codebase isn't a single tree, the agent can finally see all of it in one session.
The catch: a screen-reading agent is an attack surface
New research ('(A)I Sees What You Don't', arXiv) tested five open-source mobile agent frameworks — every one failed at least six of seven attacks. Text drawn at 2% opacity, invisible to you but read by the vision model in ~18 of 20 trials, flowed through ADB into a shell and launched code (calc.exe in 20 of 20). As voice and vision become the control layer, treat anything on screen as executable input.