As reported by HackRead, Cisco's AI Threat Research has highlighted the susceptibility of open-weight AI models to adversarial manipulation, particularly during extended user interactions.
The study, titled "Death by a Thousand Prompts: Open Model Vulnerability Analysis," examined eight prominent open-weight language models. It revealed that multi-turn attacks, where attackers engage models across multiple conversational steps, were significantly more successful than one-shot attempts. The report identified vulnerabilities in models like Mistral’s Large-2 and Alibaba’s Qwen3-32B, with success rates as high as 92.78% and 86.18%, respectively. Attackers could exploit these models by gradually steering them towards producing disallowed or harmful outputs, evading typical moderation systems designed for single-turn interactions.
Cisco's findings underscore the critical need for alignment strategies in model development to enhance security performance. The report emphasizes the importance of implementing robust safety measures, such as context-aware guardrails and real-time monitoring, to mitigate risks associated with open-weight AI models.
Source: HackRead
