Chatgpt Agent: Altman warns of use

The Chatgpt Agent can take on extensive tasks. According to Sam Altman, you shouldn’t use it extensively as a precaution. They are too great.

Chatgpt Agent is a kind of combination of the previous agent model Operator, the skills for information processing from Deep Research and the language skills of chatt. United, they should be able to represent “a new stage of the performance of AI systems” and to do “remarkable, complex tasks”, writes Altman at X. But the warning follows on the foot.

It is “a chance to try the future, but not as something that I would use for high -ranking purposes or with lots of personal information until we have the opportunity to examine and improve it in the wild.” Altman continues to try to trick the agent, Altman continues. You should therefore only give them access that the agents absolutely need to do a task – for example, access to the calendar, but not permission to buy clothing directly.

AI models are susceptible to attacks

Specifically, Altman even warns that Chatgpt Agent could reveal explosive information if you give him free access to your own emails. Someone could use a malignant mail to instruct the agent to disclose data. This could simply be written as prompt in the text field, no further knowledge is necessary. It is unclear whether the agent is really so easy to trick.

Jailbreaks and injections are generally a big problem for AI models. You are susceptible to poisoned data, hidden instructions and deliberately scattered false information. Openai emphasizes that they have taken measures for security. But apparently their effectiveness is not certain.

So what does a AI agent bring who is not allowed to act independently? Altman thinks that he can help buy an outfit for a wedding, book the trip to it and choose a gift. Just with the human supervision. Also writes Altman that Chatgpt Agent can analyze data and create a presentation. Here, too, one should limit the data it can be.