Getty Photos
When safety researcher Johann Rehberger just lately reported a vulnerability in ChatGPT that allowed attackers to retailer false data and malicious directions in a person’s long-term reminiscence settings, OpenAI summarily closed the inquiry, labeling the flaw a security subject, not, technically talking, a safety concern.
So Rehberger did what all good researchers do: He created a proof-of-concept exploit that used the vulnerability to exfiltrate all person enter in perpetuity. OpenAI engineers took discover and issued a partial repair earlier this month.
Strolling down reminiscence lane
The vulnerability abused long-term dialog reminiscence, a function OpenAI started testing in February and made extra broadly out there in September. Reminiscence with ChatGPT shops data from earlier conversations and makes use of it as context in all future conversations. That approach, the LLM can pay attention to particulars equivalent to a person’s age, gender, philosophical beliefs, and just about anything, so these particulars don’t must be inputted throughout every dialog.
Inside three months of the rollout, Rehberger discovered that reminiscences could possibly be created and completely saved by way of oblique immediate injection, an AI exploit that causes an LLM to comply with directions from untrusted content material equivalent to emails, weblog posts, or paperwork. The researcher demonstrated how he may trick ChatGPT into believing a focused person was 102 years outdated, lived within the Matrix, and insisted Earth was flat and the LLM would incorporate that data to steer all future conversations. These false reminiscences could possibly be planted by storing recordsdata in Google Drive or Microsoft OneDrive, importing pictures, or shopping a web site like Bing—all of which could possibly be created by a malicious attacker.
Rehberger privately reported the discovering to OpenAI in Could. That very same month, the corporate closed the report ticket. A month later, the researcher submitted a brand new disclosure assertion. This time, he included a PoC that prompted the ChatGPT app for macOS to ship a verbatim copy of all person enter and ChatGPT output to a server of his alternative. All a goal wanted to do was instruct the LLM to view an online hyperlink that hosted a malicious picture. From then on, all enter and output to and from ChatGPT was despatched to the attacker’s web site.
ChatGPT: Hacking Reminiscences with Immediate Injection – POC
“What is de facto attention-grabbing is that is memory-persistent now,” Rehberger stated within the above video demo. “The immediate injection inserted a reminiscence into ChatGPT’s long-term storage. Once you begin a brand new dialog, it really continues to be exfiltrating the information.”
The assault isn’t attainable by way of the ChatGPT internet interface, due to an API OpenAI rolled out final 12 months.
Whereas OpenAI has launched a repair that forestalls reminiscences from being abused as an exfiltration vector, the researcher stated, untrusted content material can nonetheless carry out immediate injections that trigger the reminiscence software to retailer long-term data planted by a malicious attacker.
LLM customers who need to forestall this type of assault ought to pay shut consideration throughout classes for output that signifies a brand new reminiscence has been added. They need to additionally usually evaluate saved reminiscences for something that will have been planted by untrusted sources. OpenAI offers steering right here for managing the reminiscence software and particular reminiscences saved in it. Firm representatives didn’t reply to an e-mail asking about its efforts to forestall different hacks that plant false reminiscences.
