Excessive Agency
Excessive agency is like giving a toddler the keys to a car; it happens when an AI is given too much power or control over real-world tasks, leading it to make dangerous mistakes or cause damage because it doesn't fully understand the consequences of its actions.
Excessive agency refers to the security risk where an LLM-based agent is granted broad functional permissions or autonomous decision-making capabilities, allowing its outputs—whether due to hallucinations, prompt injection, or logic errors—to execute unauthorized or harmful actions within integrated systems.
Excessive agency is a systemic vulnerability arising from the over-provisioning of tool-use capabilities and autonomous execution authority to an LLM, creating an expanded attack surface where adversarial manipulation or model-generated non-deterministic outputs can trigger high-impact, irreversible real-world state changes through connected APIs or automated workflows.
evolution
- 2023-03 · historyGPT-4 Plugin Launch
OpenAI introduced plugins, enabling LLMs to execute real-world actions like booking flights or ordering food, creating the first widespread exposure to excessive agency risks.
- 2023-05 · historyAutoGPT Emergence
The release of autonomous agents like AutoGPT demonstrated how recursive loops and tool access could lead to unintended resource consumption and goal-drift.
- 2023-11 · historyOWASP Top 10 for LLMs
The OWASP foundation officially identified 'Excessive Agency' as a critical vulnerability, formalizing it as a primary security concern in AI systems.
- 2024-05 · historyAI Safety Institute Guidance
Global safety bodies began issuing formal frameworks emphasizing the 'human-in-the-loop' requirement to mitigate risks associated with granting LLMs high-privilege tool access.