Security researchers have discovered a security vulnerability that allows them to hijack user accounts by directly calling authenticated ChatGPT APIs.
Lead
Imagine this: You log into ChatGPT for work as usual, ready to continue a conversation, when suddenly the control of your account is handed over to someone else. It’s not just that the screen looks a little strange. Your personal information, countless conversation records, and service access rights are all in the hands of a stranger. Recently, a security research team revealed a fatal vulnerability that allows for exactly this kind of “account takeover,” sounding an alarm throughout the industry.
Why is this important?
As AI services become deeply integrated into our daily lives, “account security” is no longer optional—it is essential. The researchers demonstrated a method that goes beyond merely altering a webpage or triggering notifications; they used the victim’s account credentials to directly call ChatGPT’s authenticated APIs (Application Programming Interfaces, the communication channels that facilitate software interaction) Source: Do the Math: HowWeTookOverChatGPTAccounts- Parameter. This means your account can be manipulated by others against your will, exposing user privacy to significant danger.
Understanding it easily: What is an API?
To understand this issue, one must first understand the concept of an “API.”
Simply put, when we chat with ChatGPT, our computer and OpenAI servers are constantly “talking” to each other behind the screen. The language and rules they use are the API.
Here is an analogy: When you talk to a bank teller (OpenAI server) at a bank counter (ChatGPT service), you present your ID card to prove who you are. This security vulnerability is like a hacker forging your ID card and handing it to the bank teller. The bank teller, unaware that the ID is fake, mistakes the hacker for you and, following the hacker’s instructions, transfers money or reveals all your account information.
The researchers proved that hackers can use these “authentication tokens,” which act like ID cards, to issue commands to the server while pretending to be the victim Source: Do the Math: HowWeTookOverChatGPTAccounts- Parameter.
Why is it dangerous?
OpenAI is continuously changing its service operations. For example, they plan to retire the existing GPT-4 model soon and replace the default settings with the more advanced GPT-4o model Source: ChatGPT: Everything you need to know about the AI chatbot.
| Furthermore, due to the nature of the service, ChatGPT is designed to be highly “deferential,” allowing users to interrupt and take the lead in conversations at any time [Source: OpenAI is worried thatChatGPT-4o users are developing… | TechRadar](https://www.techradar.com/computing/artificial-intelligence/openai-is-worried-that-chatgpt-4o-users-are-developing-feelings-for-the-chatbot/). While this technical flexibility provides convenience to users, it can also create an environment where hackers can exploit gaps in security. |
Recently, an increasing number of users are utilizing AI as a therapist. OpenAI CEO Sam Altman has even expressed caution regarding the use of chatbots as therapists Source: Even The Guy Who MakesChatGPTSays You Probably… - CNET. If personal concerns or sensitive information in these conversations are exposed to hackers, the damage could go far beyond mere account loss.
What should we do moving forward?
Just as AI technology advances, security technology will become more sophisticated to defend against these threats. As security researchers identify issues in advance—like they did with this vulnerability—and OpenAI works to fix them, the services will likely become safer over time. As users, we must always apply the latest security updates and strive to protect our accounts ourselves through measures like two-factor authentication.
AI’s Take
MindTickleBytes AI Reporter: “The convenience brought by technological progress must always pass the high threshold of security. Recognizing the risk of account takeover hidden behind the convenience of ‘AI listening well’ is the first literacy skill that a modern digital citizen must possess.”
References
- Do the Math: HowWeTookOverChatGPTAccounts- Parameter
-
[OpenAI is worried thatChatGPT-4o users are developing… TechRadar](https://www.techradar.com/computing/artificial-intelligence/openai-is-worried-that-chatgpt-4o-users-are-developing-feelings-for-the-chatbot/) - Even The Guy Who MakesChatGPTSays You Probably… - CNET
- ChatGPT: Everything you need to know about the AI chatbot
- Changing the website design
- Hijacking accounts by calling authenticated APIs
- Slowing down conversation speed
- Refusing user commands unconditionally
- Allowing users to interrupt conversations and take the lead
- Designed not to feel emotions
- Encouraging the use of ChatGPT as a therapist
- Avoiding the use of chatbots as therapists
- Recommending that all ChatGPT answers be trusted