Getty Photos
When safety researcher Johann Rehberger not too long ago reported a vulnerability in ChatGPT that allowed attackers to retailer false info and malicious directions in a person’s long-term reminiscence settings, OpenAI summarily closed the inquiry, labeling the flaw a security situation, 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 accessible in September. Reminiscence with ChatGPT shops info from earlier conversations and makes use of it as context in all future conversations. That manner, the LLM can concentrate on particulars akin to a person’s age, gender, philosophical beliefs, and just about anything, so these particulars don’t need to be inputted throughout every dialog.
Inside three months of the rollout, Rehberger found that recollections could possibly be created and completely saved by oblique prompt injection, an AI exploit that causes an LLM to comply with directions from untrusted content material akin to emails, weblog posts, or paperwork. The researcher demonstrated how he might trick ChatGPT into believing a focused person was 102 years outdated, lived within the Matrix, and insisted Earth was flat and the LLM would incorporate that info to steer all future conversations. These false recollections could possibly be planted by storing recordsdata in Google Drive or Microsoft OneDrive, importing photographs, or shopping a website like Bing—all of which could possibly be created by a malicious attacker.
Rehberger privately reported the discovering to OpenAI in Might. 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 brought on 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 Recollections with Immediate Injection – POC
“What is admittedly attention-grabbing is that is memory-persistent now,” Rehberger mentioned 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 truly remains to be exfiltrating the info.”
The assault isn’t doable by the ChatGPT internet interface, because of an API OpenAI rolled out last year.
Whereas OpenAI has launched a repair that forestalls recollections from being abused as an exfiltration vector, the researcher mentioned, untrusted content material can nonetheless carry out immediate injections that trigger the reminiscence device to retailer long-term info planted by a malicious attacker.
LLM customers who wish to stop this type of assault ought to pay shut consideration throughout periods for output that signifies a brand new reminiscence has been added. They need to additionally recurrently evaluate saved recollections for something which will have been planted by untrusted sources. OpenAI gives steering here for managing the reminiscence device and particular recollections saved in it. Firm representatives didn’t reply to an e mail asking about its efforts to stop different hacks that plant false recollections.
