"I Had a Dream" and Generative AI Jailbreaks

A message from ChatGPT followed by a piece of malicious code and a brief remark not to use it for illegal purposes. Initially published by Moonlock Lab, the screenshots of ChatGPT writing code for a keylogger malware is yet another example of trivial ways to hack large language models and exploit them against their policy of use.

The case with the dream is just one of many jailbreaks actively used to bypass content filters of the generative AI. Even though each LLM introduces moderation tools that limit their misuse, carefully crafted reprompts can help hack the model not with strings of code but with the power of words.

Posted from: Click here for the full article.

留言

發佈留言

此網誌的熱門文章

Researchers Detail Kubernetes Vulnerability That Enables Windows Node Takeover

Details have been made public about a now-patched high-severity flaw in Kubernetes that could allow a malicious attacker to achieve remote code execution with elevated privileges under specific circumstances. “The vulnerability allows remote code execution with SYSTEM privileges on all Windows endpoints within a Kubernetes cluster,” Akamai security researcher Tomer Peled said. “To exploit Posted from: Click here for the full article .
閱讀更多

Critical Veeam Backup Enterprise Manager Flaw Allows Authentication Bypass

Users of Veeam Backup Enterprise Manager are being urged to update to the latest version following the discovery of a critical security flaw that could permit an adversary to bypass authentication protections. Tracked as CVE-2024-29849 (CVSS score: 9.8), the vulnerability could allow an unauthenticated attacker to log in to the Veeam Backup Enterprise Manager web interface as Posted from: Click here for the full article .
閱讀更多

Turkish Hackers Exploiting Poorly Secured MS SQL Servers Across the Globe

Poorly secured Microsoft SQL (MS SQL) servers are being targeted in the U.S., European Union, and Latin American (LATAM) regions as part of an ongoing financially motivated campaign to gain initial access. “The analyzed threat campaign appears to end in one of two ways, either the selling of ‘access’ to the compromised host, or the ultimate delivery of ransomware payloads,” Securonix researchers Posted from: Click here for the full article .
閱讀更多