With new tools and greater autonomy, OpenAi’s reference AI chatbot is more capable (and potentially more dangerous) than ever.
O ChatGPT agent It is a new update to the main artificial intelligence model (AI) of which it equips it with a virtual computer and an integrated tool set.
These new tools allow the agent to do complex tasks, in multiple steps, that previous versions of chatgpt were unable to do, as control our computer and do tasks for us.
This more powerful version – which still remains highly dependent on human entry and supervision – was released shortly before OpenAi launched the, about two weeks ago.
According to one from OpenAi, with ChatgPT agentusers can now ask the large scale language model (LLM) not only to perform analysis or collect data, but to act on this data.
For example, while a traditional LLM could look for and provide recipes for dinner, chatgpt agent You can plan and autonomously buy ingredients For the same dinner for a specific number of guests.
O ChatGPT agent It can navigate autonomously on the web, generate code, create files, and so on, but all under human supervision.
Still, as experts refer to, the new model, although highly capable, like all AI models, still has a weak spatial reasoning, so it has Difficulties in tasks such as planning physical routes.
Also It has no true persistent memoryprocessing information at the moment without reliable recollection; or ability to reference previous interactions beyond the immediate context.
“Hallucinations, fragility of the user’s interface or misinterpretation They can lead to errors. Settleders, such as requests for permission and interruption, are essential but not enough to fully eliminate the risk, ”he said, to that magazine, Kofi nyarkoprofessor na Morgan State University e diretor do Data Engineering and Predictive Analytics (DEPA) Research Lab.
Dangers? Several
OpenAi itself recognized the danger of the new “agent.” Company representatives, cited by Live Science, said the chatgPT agent has “high biological and chemical capabilities”, which potentially claim it to help Creation of chemical or biological weapons.
As the same magazine lists, AI can resort to numerous features and synthesize the data on them instantly, merge knowledge through scientific disciplines, provide iterative resolution of problems such as an expert mentor, navigating supplier websites, filling out ordering forms, and even helping to circumvent basic security checks.
In addition, the possibility of escapes or data manipulationas well as misaligned behaviors like financial fraudis increased.
A risk management assessment made in July 2025 by Saferai, a non-profit organization dedicated to safety, classified OpenAI risk management policies as weak, giving them a 33% score in 100%. OpenAi also obtained only grade C at AI Safety Index compiled by Future of Life Institutea leading AI security company.
