A critical vulnerability known as “TokenBreak” allows attackers to bypass AI content moderation by minimally altering text. By adding a single character, malicious prompts evade detection due to differences in tokenization strategies among AI models. Models using Byte Pair Encoding and WordPiece are susceptible, while Unigram-based models remain secure, highlighting urgent security risks for organizations reliant on text classification.
India Faces Surge in FakeUpdates Malware and Remote Access Trojans: Check Point
In May, Europol, the FBI, Microsoft, and other partners targeted malware-as-a-service platform, Lumma, seizing thousands of domains. Despite the takedown, Lumma’s Russia-based servers remain operational,
