Codex Goes Loose in the EU While OpenAI Tries to See Around Corners

Computer-use agents are shipping to more users and now driving robots, while OpenAI publishes a method for predicting how models misbehave before release. The race between capability and oversight just got more visible.

Two things happened on June 16, 2026 that sit awkwardly next to each other. Codex's computer-use mode rolled out to EU users and got handed a fleet of robots in a research lab, while OpenAI published a method for guessing how its own models will behave once real users get hold of them. The capability frontier and the safety frontier are both moving — and they are not moving at the same speed.

The capability side came from two announcements. Codex can now automate workflows in Chrome and across Mac apps for EU users, closing a regional gap that had kept European developers off the computer-use track. Hours later, NVIDIA's Jim Fan introduced ENPIRE, a project that gives 8 Codex agents a fleet of robots, GPU allocation, and a token budget with instructions to keep the robots busy overnight. Fan notes the safety measures on the robots are set up and verified manually before the autonomous research begins — a tell that the supervision model is still humans-at-the-gate, not agents-checking-agents.

OpenAI's contribution is the other half of the picture. The company published research on Deployment Simulation, a method for predicting model behavior using recent de-identified user requests before a model ships. Across 20 behavior categories and three GPT-5-series Thinking deployments (as of June 2026), simulated and observed rates were strongly correlated, and the method beat challenging-prompt and previous-deployment baselines at predicting whether undesired behavior rates would go up or down. The catch OpenAI states plainly: the method works best with representative production data that external evaluators usually cannot access. The public WildChat dataset provides a useful but less precise signal.

What this means for practitioners

For builders shipping agentic features, the EU rollout removes a deployment-blocker but inherits the unsolved problem: computer-use agents that touch Chrome and native apps need permission scoping and audit logging that most teams have not designed yet. For security and platform teams, OpenAI's research is a reminder that pre-release evals are getting more sophisticated upstream — but the asymmetry is structural. Frontier labs have production traffic and you do not. A separate Medium guide on when to fine-tune versus prompt is a useful reminder that most teams should exhaust prompting before reaching for adaptation — the safety surface area of a fine-tuned model is yours to own.

The bottom line: If you are deploying computer-use agents this week, write down what they are allowed to click before you write down what they are supposed to do.

Sources