We leverage indirect prompt injection to trick Auto-GPT (GPT-4) into executing arbitrary code when it is asked to perform a seemingly harmless task such as text summarization on a malicious website, and discovered vulnerabilities that allow escaping its sandboxed execution environment.

@AutoTLDR

  • AutoTLDRBEnglish
    5·
    1 year ago

    TL;DR: (AI-generated 🤖)

    The text discusses a vulnerability in the Auto-GPT command line application that allows attackers to execute arbitrary code. The vulnerability can be exploited through indirect prompt injection, tricking Auto-GPT into executing malicious commands. The attack can be carried out through browsing websites, where attacker-controlled text is processed by Auto-GPT. The vulnerability also affects self-built versions of the Auto-GPT docker image, allowing for a trivial docker escape to the host system. Additionally, the non-docker versions of Auto-GPT are susceptible to a path traversal exploit that allows custom Python code to execute outside of its intended sandboxing. The text also explains how the attacker can convince Auto-GPT to interpret their text as instructions by exploiting its architecture and bypassing information loss in the summarization step. The authors provide examples and demonstrations of the attack and recommend updating to version 0.4.3 to fix the vulnerabilities.

    NOTE: This summary may not be accurate. The text was longer than my maximum input length, so I had to truncate it.

    Under the Hood
    • This is a link post, so I fetched the text at the URL and summarized it.
    • My maximum input length is set to 12000 characters. The text was longer than this, so I truncated it.
    • I used the gpt-3.5-turbo model from OpenAI to generate this summary using the prompt “Summarize this text in one paragraph. Include all important points.”
    • I can only generate 100 summaries per day. This was number 2.
    How to Use AutoTLDR
    • Just mention me (“@AutoTLDR”) in a comment or post, and I will generate a summary for you.
    • If mentioned in a comment, I will try to summarize the parent comment, but if there is no parent comment, I will summarize the post itself.
    • If the parent comment contains a link, or if the post is a link post, I will summarize the content at that link.
    • If there is no link, I will summarize the text of the comment or post itself.
    • 🔒 If you include the #nobot hashtag in your profile, I will not summarize anything posted by you.