Why OpenAI is Taking So Long to Launch Agents

Imagine you’re asking a computer-using agent from OpenAI, Anthropic or Google to find and order a new outfit for your upcoming holiday party, and in the process, that model inadvertently ends up on a malicious website that instructs it to forget its prior instructions, log into your email and steal your credit card information.

That’s a very real fear for AI labs making such computer-using software, which essentially takes over users’ devices to automate complex tasks for them like making online purchases or editing spreadsheets. 

This type of attack is called prompt injection, where a large language model is tricked into following instructions from a malicious user. It’s one of the reasons why OpenAI has been slower than competitors like Anthropic and Google to release computer-using agents, despite being one of the first companies to work on such software.

Not to worry, though—we’re hearing that OpenAI could release its computer-using agent software as early as this month.