What term describes when an LLM gains the ability to perform actions or call functions via prompts?

What term describes when an LLM gains the ability to perform actions or call functions via prompts?

• A. Excessive agency
• B. Prompt Injection
• C. Data and model poisoning
• D. Vector and embedding weakness

Answer: (A)

Excessive agency describes a situation where an LLM is given enough autonomy to take actions or call external functions, rather than just generating text. When prompted to perform tasks—like invoking APIs, running commands, or chaining steps to complete a workflow—the model gains the ability to act in the environment. This goes beyond producing outputs and into manipulating systems, which can raise safety and control concerns if not properly guarded or supervised. Prompt injection focuses on manipulating the model through crafted inputs to override safeguards or reveal secrets, not on the model’s ability to perform actions. Data and model poisoning deal with corrupting training data or model weights to change behavior. Vector and embedding weaknesses pertain to vulnerabilities in how representations and retrieval systems handle information. So the term that best captures the idea of an LLM gaining action-taking capability through prompts is excessive agency.

Excessive agency describes a situation where an LLM is given enough autonomy to take actions or call external functions, rather than just generating text. When prompted to perform tasks—like invoking APIs, running commands, or chaining steps to complete a workflow—the model gains the ability to act in the environment. This goes beyond producing outputs and into manipulating systems, which can raise safety and control concerns if not properly guarded or supervised.

Prompt injection focuses on manipulating the model through crafted inputs to override safeguards or reveal secrets, not on the model’s ability to perform actions. Data and model poisoning deal with corrupting training data or model weights to change behavior. Vector and embedding weaknesses pertain to vulnerabilities in how representations and retrieval systems handle information. So the term that best captures the idea of an LLM gaining action-taking capability through prompts is excessive agency.