The most recent generative AI fashions aren’t simply stand-alone text-generating chatbots—as a substitute, they will simply be hooked as much as your information to offer customized solutions to your questions. OpenAI’s ChatGPT can be linked to your Gmail inbox, allowed to examine your GitHub code, or discover appointments in your Microsoft calendar. However these connections have the potential to be abused—and researchers have proven it may possibly take only a single “poisoned” doc to take action.
New findings from safety researchers Michael Bargury and Tamir Ishay Sharbat, revealed on the Black Hat hacker convention in Las Vegas at present, present how a weak point in OpenAI’s Connectors allowed delicate info to be extracted from a Google Drive account utilizing an indirect prompt injection attack. In an illustration of the assault, dubbed AgentFlayer, Bargury exhibits the way it was potential to extract developer secrets and techniques, within the type of API keys, that had been saved in an illustration Drive account.
The vulnerability highlights how connecting AI fashions to exterior methods and sharing extra information throughout them will increase the potential assault floor for malicious hackers and probably multiplies the methods the place vulnerabilities could also be launched.
“There’s nothing the person must do to be compromised, and there’s nothing the person must do for the information to exit,” Bargury, the CTO at safety agency Zenity, tells WIRED. “We’ve proven that is fully zero-click; we simply want your electronic mail, we share the doc with you, and that’s it. So sure, that is very, very dangerous,” Bargury says.
OpenAI didn’t instantly reply to WIRED’s request for remark concerning the vulnerability in Connectors. The corporate launched Connectors for ChatGPT as a beta function earlier this 12 months, and its website lists no less than 17 totally different companies that may be linked up with its accounts. It says the system permits you to “deliver your instruments and information into ChatGPT” and “search recordsdata, pull dwell information, and reference content material proper within the chat.”
Bargury says he reported the findings to OpenAI earlier this 12 months and that the corporate shortly launched mitigations to forestall the approach he used to extract information through Connectors. The way in which the assault works means solely a restricted quantity of knowledge may very well be extracted directly—full paperwork couldn’t be eliminated as a part of the assault.
“Whereas this subject isn’t particular to Google, it illustrates why growing sturdy protections towards immediate injection assaults is necessary,” says Andy Wen, senior director of safety product administration at Google Workspace, pointing to the corporate’s recently enhanced AI security measures.
