The new feature allows ChatGPT to operate as its own virtual computer, enabling it to independently complete complex, multi-step tasks. Photo: Emiliano Vittoriosi/Unsplash
“>
The new feature allows ChatGPT to operate as its own virtual computer, enabling it to independently complete complex, multi-step tasks. Photo: Emiliano Vittoriosi/Unsplash
OpenAI has announced a new upgrade to ChatGPT, introducing “agentic capabilities” that enable the AI model to perform complex tasks autonomously while possessing what the company describes as “high biological and chemical capabilities”.
The new feature allows ChatGPT to operate as its own virtual computer, enabling it to independently complete complex, multi-step tasks – from web research to document creation – through what OpenAI calls an “agentic” system that combines browsing, analysis, and execution capabilities.
For all latest news, follow The Daily Star’s Google News channel. Due to these expanded functions, OpenAI has classified the new agent system as having “high biological and chemical capabilities” under its preparedness framework. While the company states there’s no evidence the AI could meaningfully assist in creating biological threats, it has implemented enhanced safety measures as a precaution.
These measures include specialised training to refuse dual-use requests, continuous monitoring systems, and collaboration with biosecurity experts to assess risks. According to OpenAI, the new feature will actively block attempts to obtain dangerous biological information and will require human confirmation for sensitive actions.
Users can activate the agent mode through a dropdown menu, with free-form conversation transitioning into task execution. Current limitations include basic slide deck formatting and occasional export inconsistencies, which OpenAI says will improve in future updates.
Security measures include optional connector services for app integration, private browsing sessions that don’t store login data, and one-click data deletion. However, OpenAI acknowledges increased risks from the AI’s ability to take real-world actions, particularly regarding potential manipulation through hidden web prompts.
OpenAI also says that users can retain control through pause and intervention options during task execution. Additionally, the company has established a bug bounty program to identify vulnerabilities as this powerful new functionality becomes widely available.
The feature is rolling out first to ChatGPT Pro users, followed by Plus and Team subscribers in the coming days, with Enterprise and Education access planned for the weeks ahead, according to OpenAI.
