LLMs can be exploited by attackers using crafted inputs, leading to unintended execution of their commands, potentially resulting in data breaches and social engineering attacks.

Training data poisoning compromises model security, effectiveness, and ethical behavior, posing risks of performance degradation and reputational damage.

LLM applications may unintentionally reveal sensitive data, risking unauthorized access and intellectual property theft, necessitating robust privacy safeguards.

Insecure Output Handling in LLMs can lead to XSS, CSRF, SSRF, privilege escalation, and remote code execution in downstream components

Insufficient access controls and input validation in plugins can lead to data exfiltration, remote code execution, and privilege escalation, highlighting the need for robust security measures.

Overly permissive functionality and autonomy in LLM-based systems pose vulnerability risks, urging developers to restrict plugin capabilities, permissions, and autonomy to essential levels, enforce user authorization, mandate human approval for all actions, and integrate authorization in downstream systems.

Unauthorized access and exfiltration of LLM models pose risks of economic loss, reputation damage, and unauthorized data access, necessitating robust security measures.

Model Denial of Service involves attackers interacting with an LLM, consuming excessive resources, leading to degraded service quality and increased resource costs.