Artificial Intelligence & Machine Learning
,
Next-Generation Technologies & Secure Development
Operator Cannot Yet Reliably Perform Complex, Customized Tasks
OpenAI introduced an AI agent capable of independent action with the launch of Operator, a general-purpose AI tool that interacts with websites to perform tasks.
See Also: Live Webinar | AI in the Spotlight: Exploring the Future of AppSec Evolution
The AI agent can navigate menus and complete forms to do tasks such as travel booking, ordering takeout, buying stuff or scheduling tasks.
OpenAI CEO Sam Altman called Operator a basic step toward unlocking the power of AI agents, previously describing 2025 as the year that would be decisive for this technology. While streaming live, Altman announced that plans were underway to extend Operator’s reach to other countries, but Europe would lag due to regulatory complexities.
The AI agent operates through a computer-using agent model, enabling the agent to interact with website frontends without requiring developer-facing APIs. Operator can press buttons, navigate drop-down menus and input information into forms.
In the initial phase of implementation, OpenAI is collaborating with companies including DoorDash, eBay, Instacart, Priceline, StubHub and Uber. The agent is programmed to request confirmation from the user before finalizing an action, such as submitting an order or an email, to allow users to review the agent’s work.
The model has “already proven useful in a variety of cases, and we aim to extend that reliability across a wider range of tasks,” OpenAI said.
The company also acknowledged limitations, saying that Operator cannot reliably perform complex or customized tasks, such as creating intricate presentations or navigating non-standard interfaces.
For sensitive tasks, such as monetary transactions, users must manually enter information such as payment card details. They must also actively supervise the agent’s actions on sensitive websites such as email platforms.
Its inbuilt protection against abuse includes a monitoring system that terminates the agent’s activity when it notices behavior it deems suspicious, as well as automated and human-reviewed pipelines continually update protection mechanisms.
Operator also needs user intervention for CAPTCHA checking or to navigate complex web interfaces. OpenAI imposed rate limits on the number of tasks that Operator can complete in a day or at one time. Sending emails or deleting calendar events are currently unavailable because of security reasons.
Giving AI even this level of control comes with significant security considerations, said Alon Levin, product manager for Seraphic Security. This includes potential account compromise and data exfiltration due to unauthorized AI actions, unintentional access and unintended interaction with malicious sites, he told Information Security Media Group. AI agents introduce safety-related risks such as misuse for phishing scams or automated ticket scalping (see: Claude’s Computer Use May End Up a Cautionary Tale).
Operator is initially available as a research preview to U.S. users subscribed to ChatGPT’s Pro plan. OpenAI expects to eventually incorporate Operator into all ChatGPT applications.
OpenAI said the agent is safe enough to release as a research preview. “Operator makes use of tools that are designed to help limit the model’s susceptibility to malicious prompts, hidden instructions and phishing attempts,” the company said. The approach follows similar initiatives, such as Google’s Project Mariner, which also avoids automating high-risk tasks such as entering credit card numbers.
Operator’s launch follows the recent introduction of Tasks, a feature that allows ChatGPT users to automate simple activities such as setting reminders and scheduling prompts. While Tasks made ChatGPT more comparable to existing virtual assistants such as Siri and Alexa, Operator represents a more ambitious leap toward autonomous AI.