Cisco’s State of AI Security report found that open-weight AI models are highly vulnerable to multi-turn jailbreak attacks, with a 92.78% success rate. These attacks, which use iterative prompts to bypass content filters, highlight the need for improved AI security measures. The report also emphasizes the risks associated with excessive agency in AI systems, particularly when they are granted broad autonomous authority over tools and data.
https://www.databreachtoday.com/open-weight-ai-models-fail-jailbreak-test-a-30823