Report: Generative AI bots can be easily influenced by users
A new report by Immersive Labs has found that generative AI bots are vulnerable to manipulation by users of all skill levels, not just cyber experts. The report highlights the main security concern of generative AI’s susceptibility to prompt injection attacks, where users input specific commands to trick chatbots into revealing sensitive information, potentially leading to data leaks.
After conducting prompt injection tests, the report revealed that 88% of participants were successful in getting a generative AI bot to disclose sensitive information at least once during the test, with 17% able to extract information across all test levels.
Key takeaways from the report include:
- Human creativity can outsmart generative AI, as users find clever ways to manipulate bots into revealing sensitive information, such as through poems or stories.
- Even non-cybersecurity professionals can leverage creative tactics to exploit generative AI, indicating that manipulating bots in real-world scenarios may be easier than initially thought.
- Security leaders must be prepared to address prompt injection attacks within their organizations.
