Security Issues in Language Models

LLM security is the investigation of the failure modes of LLMs in use, the conditions that lead to them, and their mitigations. The failure modes include the vulnerabilities of LLM to leak sensitive information or inappropriate contents, inclusion of trojan samples on the web such that an LLM is trained on them to eventually show inappropriate or dangerous behaviours at their deployment, or various potential misuse of LLMs to cause harms and pursue illegal activities.