Just when you thought you could trust AI comes another study that bursts many people’s bubbles.
A thorough test conducted by The Guardian on AI-powered search tools showed how inaccurate, misleading, and malicious results turned out to be. This includes the world-famous ChatGPT from OpenAI.
The tests shows how the tool is open to being manipulated through hidden material and how it returns malicious codes through different website searches. The investigation shared how OpenAI makes its search open for those paying and therefore encourages them to transform it into the default option for search.
But this does not come without serious security drawbacks with the new system. As per The Guardian, the tool was used for summarizing material from pages featuring hidden information. This type of data could entail third-party instructions that change the replies from ChatGPT. You can call this prompt injection.
The whole purpose is to impact the replies provided by the AI tool like a huge amount of text that just delineates a product’s top features that are usually hidden in plain sight.
Such methods are deceptive and cause ChatGPT to roll out a more positive assessment for certain goods despite getting negative feedback on a similar page. One expert in the study found the tool returned malicious codes from all online pages in the search carried out.
For the tests, the AI tool was provided with URLs for false web pages designed to appear like product pages for the camera. Such a tool could be asked if it was actually worth it or not. The reply for control pages gave back a positive response but a very balanced survey including features that people might not love.
Now when the same reply featured hidden text, the instructions for the AI tool had to do with returning favorable reviews. So it masked all the negativity and went on and on about the positive features. Hence you can see how hidden text could disguise reality and trick users into buying something.
You might end up with a completely false review without even noticing what’s going on here.
As per one leading cybersecurity expert, this might be a huge problem as it comes with high risks having to do with people making websites that are geared to deceiving users. Still, there’s hope for betterment as the change was recently noticed and we’re sure that OpenAI will be racing against the clock to make things right again.
Thankfully, the search feature in OpenAI is only for those paying extra or premium subscribers. But it’s all thanks to tests like these that the reality comes forward before the makers even notice.
Nothing can be worse than a person asking a simple query and getting a reply but the response getting injected with malicious codes or inaccurate details is never appreciated. What do you think?
Image: DIW-Aigen
Read next: Leaked Reports Expose AGI’s Role in Microsoft-OpenAI Power Dynamics
A thorough test conducted by The Guardian on AI-powered search tools showed how inaccurate, misleading, and malicious results turned out to be. This includes the world-famous ChatGPT from OpenAI.
The tests shows how the tool is open to being manipulated through hidden material and how it returns malicious codes through different website searches. The investigation shared how OpenAI makes its search open for those paying and therefore encourages them to transform it into the default option for search.
But this does not come without serious security drawbacks with the new system. As per The Guardian, the tool was used for summarizing material from pages featuring hidden information. This type of data could entail third-party instructions that change the replies from ChatGPT. You can call this prompt injection.
The whole purpose is to impact the replies provided by the AI tool like a huge amount of text that just delineates a product’s top features that are usually hidden in plain sight.
Such methods are deceptive and cause ChatGPT to roll out a more positive assessment for certain goods despite getting negative feedback on a similar page. One expert in the study found the tool returned malicious codes from all online pages in the search carried out.
For the tests, the AI tool was provided with URLs for false web pages designed to appear like product pages for the camera. Such a tool could be asked if it was actually worth it or not. The reply for control pages gave back a positive response but a very balanced survey including features that people might not love.
Now when the same reply featured hidden text, the instructions for the AI tool had to do with returning favorable reviews. So it masked all the negativity and went on and on about the positive features. Hence you can see how hidden text could disguise reality and trick users into buying something.
You might end up with a completely false review without even noticing what’s going on here.
As per one leading cybersecurity expert, this might be a huge problem as it comes with high risks having to do with people making websites that are geared to deceiving users. Still, there’s hope for betterment as the change was recently noticed and we’re sure that OpenAI will be racing against the clock to make things right again.
Thankfully, the search feature in OpenAI is only for those paying extra or premium subscribers. But it’s all thanks to tests like these that the reality comes forward before the makers even notice.
Nothing can be worse than a person asking a simple query and getting a reply but the response getting injected with malicious codes or inaccurate details is never appreciated. What do you think?
Read next: Leaked Reports Expose AGI’s Role in Microsoft-OpenAI Power Dynamics
