7 ChatGPT Security Vulnerabilities That Are Vulnerable to Data Infiltration
Research reveals security vulnerabilities in chatgpt that could potentially leak user data
Cyber security researchers from Tenable Research found a number of security holes in the chatgpt system that made it vulnerable to data theft. In the report entitled HackedGpt: Novel Ai Vulnerabilities Open The Door For Private Data Leakage, it was found that this gap allows the attacker to steal information from chatgpt memory and user’s conversation history without realizing it. The study also shows the vulnerability that arises from the way chatgpt processing inputs and the context of the conversation.
Here are seven types of vulnerability or security vulnerabilities in chatgpt that have the potential to leak user personal data based on the findings of Tenable Research researchers:
Indirect Prompt Injection
Attacks are inserted through comments or hidden content on the website that the user asks to summarize. So that when the chatgpt reads the accessed site, malicious instructions are also executed.Zero-Click Prompt Injection
When the user asks a simple question, chatgpt then explores the malicious sites that have been indexed, without the need for additional clicks.One-Click Prompt Injection
The attacker can create a malicious link that doesn’t look suspicious, like https://chatgpt.com/?q={Prompt}.#ATFP_CLOSE_TRANSLATE_SPAN# When a user clicks on the link, ChatGPT automatically runs preset malicious commands.Bypass security mechanism
Features url_safe Owned OpenAI should have a primary function to block malicious links. However, researchers found a gap through the domain bing.com/ck/a which can be used to bypass the filtering system and steal letter by letter data.Conversation Injection
Through this technique, SearchGPT can write hidden instructions in the context of a conversation, then chatGPT reads it as if it came from the user. As a result, ChatGPT actually injects itself with malicious commands.Malicious Content Hiding
Dangerous instructions can be hidden inside Code block markdown (How to display code or text that has been formatted in a markdown document). The danger is, the user does not see the hidden text, but the chatgpt still reads it and runs the command.Memory injection
An attacker can force chatgpt to update its memory to always run certain instructions. This attack can be very dangerous because it is persistent and can take place even though the conversation session is over.
The researchers assess, there are hundreds of millions of potentially affected chatgpt users because they use this model every day for various needs. Attacks can cause leakage of personal data stored in the Memory feature; Bypass chatgpt security system; and continuous attacks that remain active across conversation sessions.
ChatGPT is said to have access to various internal tools, including Web Tool to browse the internet, Search Command for search, and Memory Feature to remember user information. When these features interact with external sources such as websites or online documents, the attacker can insert hidden instructions (Injection prompt). As a result, ChatGPT can run unwanted commands and send user personal data to third-party servers.
Tenable Research researchers have also revealed all these issues to OpenAI and are working together directly to fix some of these security vulnerabilities.
