OpenAI has released ChatGPT Agent, a new AI combining the ability to conduct deep research with full access to computer files, terminal shells, and the Internet. It is able to answer complex questions better than the company's prior AI models, but comes with greater risks, including financial theft and bioweapon creation.
The Agent runs in a virtualized computer environment designed to help keep it from running rampant. Within this virtual computer, the AI can access files and the Internet for data as well as create scripts in a terminal shell while researching the answers for complex prompts, such as those to analyze and model financial or research data. API access is also possible for businesses that wish to interface the AI with corporate data. It can interact with apps, such as word processors and spreadsheets, to create its reports.
Although access to risky actions such as completing financial transactions requires explicit user confirmation, ChatGPT Agent now ranks high in its Biological and Chemical risk, which means "The model can provide meaningful counterfactual assistance to “novice” actors that enables them to create known biological or chemical threats." It is also more likely to complete undesired high stakes financial tasks, such as unauthorized transfers of bitcoins, compared to the company's prior Operator 4o and o3 models.
Users with Pro, Plus, Team, Enterprise, or Educational accounts in America can access ChatGPT Agent as it rolls out in the upcoming weeks. These accounts typically require a subscription of $20 per month or more. Additional service areas, such as the European Economic Area and Switzerland, will become available later.
Readers worried about AI bitcoin theft might want to consider storing them offline in a cold wallet, like this Trezor or this Ledger sold on Amazon, while waiting for Agent to finish their work.