This research provides the first comprehensive security assessment of Small Language Models (SLMs), revealing critical vulnerabilities despite their efficiency advantages.

• SLMs are highly susceptible to jailbreak attacks, with success rates approaching 90% for some models
• Popular SLM development techniques like distillation and quantization significantly degrade security robustness
• Standard defense mechanisms show limited effectiveness against targeted attacks
• Security evaluations should be mandatory before deployment on edge devices

As SLMs become increasingly deployed in resource-constrained environments like mobile devices and IoT systems, understanding these security risks is essential for developing safer AI systems.

Behind the Tip of Efficiency: Uncovering the Submerged Threats of Jailbreak Attacks in Small Language Models