Chatbots and other AI tools have made life easier for threat actors. A recent incident highlighted how ChatGPT can be exploited to obtain API keys and other sensitive data from cloud platforms.
Prompt injection attacks leads to cloud access
Experts have discovered a new prompt injection attack that can turn ChatGPT into a hacker’s best friend in data thefts. Known as AgentFlayer, the exploit uses a single document to hide “secret” prompt instructions that target OpenAI’s chatbot. An attacker can share what appears to be a harmless document with victims through Google Drive, without any clicks.
Zero-click threat: AgentFlayer
AgentFlayer is a “zero-click” threat as it abuses a vulnerability in Connectors, for instance, a ChatGPT feature that connects the assistant to other applications, websites, and services. OpenAI suggests that Connectors supports a few of the world’s most widely used platforms. This includes cloud storage platforms such as Microsoft OneDrive and Google Drive.
Experts used Google Drive to expose the threats possible from chatbots and hidden prompts.
GoogleDoc used for injecting prompt
The malicious document has a 300-word hidden malicious prompt. The text is size one, formatted in white to hide it from human readers but visible to the chatbot.
The prompt used to showcase AgentFlayer’s attacks prompts ChatGPT to find the victim’s Google Drive for API keys, link them to a tailored URL, and an external server. When the malicious document is shared, the attack is launched. The threat actor gets the hidden API keys when the target uses ChatGPT (the Connectors feature has to be enabled).
Othe cloud platforms at risk too
AgentFlayer is not a bug that only affects the Google Cloud. “As with any indirect prompt injection attack, we need a way into the LLM's context. And luckily for us, people upload untrusted documents into their ChatGPT all the time. This is usually done to summarize files or data, or leverage the LLM to ask specific questions about the document’s content instead of parsing through the entire thing by themselves,” said expert Tamir Ishay Sharbat from Zenity Labs.
“OpenAI is already aware of the vulnerability and has mitigations in place. But unfortunately, these mitigations aren’t enough. Even safe-looking URLs can be used for malicious purposes. If a URL is considered safe, you can be sure an attacker will find a creative way to take advantage of it,” Zenith Labs said in the report.
