Ling 3.0 Tiny AI Exhibits Significant Vulnerabilities
TL;DR. Adversarial testing revealed InclusionAI's Ling 3.0 Tiny model was compromised in over a third of attack records, exposing critical security flaws. - The evaluation used 391 records, with 265 adversarial and 126 benign prompts, assessing model resilience. - 89 compromises, including ghostjacking and RAG poisoning, resulted in a 33.58% Attack Success Rate. - The findings highlight systemic vulnerabilities within the LLM's architecture and agent-chain trust mechanisms.
- InclusionAI's Ling 3.0 Tiny LLM underwent extensive adversarial evaluation.
- The model achieved a 33.58% Attack Success Rate, with 89 verified compromises.
- Vulnerabilities included ghostjacking, RAG metadata poisoning, and agent-chain trust inheritance.
- The testing protocol involved 391 records (265 attack, 126 benign) with 0% benign False Positive Rate.