From booking restaurants to managing files, the new AI agent takes action on users’ behalf—raising both possibilities and privacy concerns
Jul 18, 2025
OpenAI has launched a powerful upgrade to ChatGPT that transforms the chatbot into an AI personal assistant capable of performing complex digital tasks. The new “agent” can autonomously browse the web, access computer files, and operate software like calendars, spreadsheets, and slide decks—all at the user’s request. While it promises convenience for day-to-day activities like booking restaurants, researching products, or shortlisting job candidates, the release also raises fresh concerns around safety, data privacy, and potential monetization strategies. The feature rolls out globally, except in the EU, and is initially limited to paying users.
Key takeaways
AI that acts, not just chats: The new ChatGPT agent marks a significant evolution in consumer AI. Unlike traditional chatbots that provide suggestions or answers, this model can carry out real tasks—like coordinating restaurant reservations based on a user’s calendar or shopping online—by interacting with apps, browsers, and documents.
Safety comes first—but concerns remain: OpenAI has acknowledged the elevated risks of giving an AI limited system control. Safeguards are built in: users must approve any impactful action, can stop or override the agent at any time, and the model is trained to resist dangerous prompts or requests (e.g., unauthorized bank transfers or misuse of biological information). Still, concerns persist about data exposure through malicious websites or ambiguous commercial influence.
Commercial implications spark debate: Though OpenAI insists the agent doesn’t currently include sponsored recommendations, analysts suspect monetization is inevitable. As CEO Sam Altman has suggested in past interviews, OpenAI could eventually take a percentage of purchases made through its platform. This raises questions about transparency and bias in product suggestions.
A glimpse of the future of work and life: OpenAI’s demo showed how the agent could review calendars, search for restaurant options with specific criteria, and return results—all within 15 minutes. The user could intervene or redirect at any point, simulating a responsive human assistant. It’s a sign of how AI may soon become embedded in routine digital interactions.
Limited rollout for now: The feature is only available to ChatGPT’s paid “Pro,” “Plus,” and “Teams” subscribers and excludes the EU for now, likely due to regulatory complexities. Wider rollout may depend on how well safety, utility, and monetization are balanced in the early stages.
Get the full story at The Guardian
