OpenAI Unveils ChatGPT agent, Bringing Agentic Capabilities to the Masses
OpenAI is unveiling ChatGPT agent, a series of new agentic capabilities that allow ChatGPT to handle complex, multi-step tasks executed by its own virtual computer. Available for Pro, Plus, and Team users, ChatGPT’s agentic features seamlessly switch between reasoning and action, handling tasks end-to-end based on user instructions.
These new agentic capabilities are underscored by a unified agentic system, which builds on the strengths of OpenAI’s previous innovations: Operator’s ability to interface with websites, deep research’s information synthesis, and ChatGPT’s intelligence and conversational fluency. OpenAI describes ChatGPT agent as a “natural evolution” of Operator and deep research, integrating each solution’s strengths into a single model that can actively engage with websites as well as execute in-depth analysis.
More granularly, ChatGPT agent consists of a suite of tools including a visual browser that interacts with the web through a graphical-user interface, a text-based browser for simpler reasoning-based web queries, a terminal, and direct API access. Additionally, the agent can utilize ChatGPT connectors to connect to apps such as Gmail and Github for integrating relevant information.
As an example, ChatGPT agent “can gather information about your calendar through an API, efficiently reason over large amounts of text using the text-based browser, while also having the ability to interact visually with websites designed primarily for humans,” explained OpenAI.
Underpinning ChatGPT agent is a dedicated virtual computer, engineered to preserve task context even when multiple tools are being used at once. Adapting its approach with speed, accuracy, and efficiency in mind, ChatGPT agent can “choose to open a page using the text browser or visual browser, download a file from the web, manipulate it by running a command in the terminal, and then view the output back in the visual browser,” according to OpenAI.
However, with broader, more powerful capabilities comes new opportunities for risk, OpenAI noted. Since ChatGPT agent can take actions on the web while working with your unique data via connectors and other websites, OpenAI has taken extensive measures to bring awareness to the potential risks as well as mitigate harm.
“We’ve strengthened the robust controls from Operator’s research preview and added safeguards for challenges such as handling sensitive information on the live web, broader user reach, and (limited) terminal network access,” said OpenAI. “We’ve placed a particular emphasis on safeguarding ChatGPT agent against adversarial manipulation through prompt injection, which is a risk for agentic systems generally, and have prepared more extensive mitigations accordingly.”
Some of OpenAI’s latest safeguards include:
- Explicit user confirmation, where ChatGPT explicitly asks for permission before taking actions with “real-world consequences,” such as making a purchase
- Active supervision (“Watch Mode”) that requires active oversight for certain critical tasks, such as sending emails
- Proactive risk mitigation where ChatGPT is trained to refuse high-risk tasks, such as bank transfers
- Privacy controls, where—through a single click in ChatGPT’s settings—users can delete all browsing data and immediately log out of all active website sessions (otherwise, cookies will persist based on each website’s individual cookie policies)
- Secure browser “Takeover Mode” that privatizes inputs (ChatGPT does not collect or store any data during these sessions)
- Protection against biological risk such as comprehensive threat modeling, dual-use refusal training, always-on classifiers and reasoning monitors, and clear enforcement pipelines (while OpenAI has no evidence to suggest ChatGPT agent can be used to create severe biological harm, the company has implemented safeguards as a measure of caution)
Furthermore, OpenAI noted that, “ChatGPT agent is still in its early stages. It’s capable of taking on a range of complex tasks, but it can still make mistakes…Overall, we expect continued improvements to ChatGPT agent’s efficiency, depth, and versatility over time, including more seamless interactions as we continue to adjust the amount of oversight required from the user to make it more useful while ensuring it’s safe to use.”
To learn more about ChatGPT agent, please visit https://openai.com/.