AI browsers have only just stepped into the real world but security issues have already sparked concerns among those in the industry. Experts have loudly talked about the impact of letting AI do your tasks
and now OpenAI is sharing its own worries and even highlighted how it plans to fight the threat.
The company plans to use AI to fight the risks posed by prompt injections with AI browsers and allow the model to train on the tricks and loopholes that the hackers can exploit.
OpenAI Ready To Fight Prompt Injections Threat: Here’s How
Before we talk about OpenAI’s plan to fight the threat, it is important to understand what prompt injection attacks are. We all know that AI chatbots respond to prompts, be it for creating an image, ordering groceries or even asking the AI to scan through a document.
For instance, when you make an AI shop for your grocery, it is likely to include the payment details as well, and these bad actors could use prompt injections to make the AI model reveal those numbers. These injections can also attack users who simply get the AI to summarise a malicious web page, it is that simple.
Now let’s talk about what OpenAI is intending to do to prevent these attacks. The company has built its own AI attacker that can disguise and simulate new prompt injections on the go. The tool is put to test on ChatGPT Atlas browser which allows them to analyse the issues with the AI browser and how they can be stopped.
The company feels that prompt injections will never be fully solved, which is never a good thing to hear but it plans to tackle the risks by building a defensive layer that can preempt and even identify the issues faster.
It is quite clear that OpenAI is not assuring users about the risks posed by these attacks on its AI browser but any long term fix for these risks will help in a big way, not only for Atlas but other similar browsers.
The company mentioned that prompt injections could be as common as online scams and stopping them might be hard but preventing and being cautious about them could be the way to go.
