Why Local LLMs Are Vulnerable to Prompt Attacks: Qwen3.6 vs Llama3.1 Security Showdown

This security evaluation reveals critical vulnerabilities in open-source LLMs when subjected to adversarial prompting, with Qwen3.6 generating attack content for 73% of requests and Llama3.1 for 33%—both falling victim to education-framed attacks. The findings demonstrate that CTF-disguised prompts effectively bypass safety mechanisms, highlighting why external security classifiers are essential rather than relying on inherent model safeguards in production environments.