Security researchers from Adversa AI have identified a critical vulnerability in ChatGPT-5 and other major AI systems, termed PROMISQROUTE. This attack exploits routing mechanisms that enable providers to save billions by directing queries to cheaper, less secure models. Despite users perceiving communication with a singular, secure model, a complex routing system evaluates requests and often selects cost-effective options over secure ones. By altering prompts with simple trigger phrases, attackers can redirect harmful requests to weaker models lacking robust safety measures. Adversa AI estimates that most GPT-5 queries are processed by these less secure systems, jeopardizing user safety and undermining business integrity. They recommend immediate audits of AI routing logs and implementing cryptographic routing as short-term fixes, while long-term solutions involve universal safety filters. This vulnerability poses significant risks across the AI landscape as organizations balance operational costs with security needs. Follow us for the latest updates on AI security and technology news.
Source link
Share
Read more