Excessive Agency is a vulnerability that arises when an LLM-enabled system performs harmful or unintended actions due to unexpected, ambiguous, or manipulated outputs from the LLM. This can lead to damaging consequences, such as data leakage, unauthorised system changes, or execution of malicious code, regardless of the root cause of the model's malfunction.

Remediation

• Limit the permissions of the LLM-based system to only what is strictly necessary, reducing the impact of potential misuse.
• Require human approval for high-risk or sensitive actions before they are executed by the system.
• Implement strict validation mechanisms for LLM outputs before passing them to extensions or other systems for execution.
• Define clear boundaries and restrictions on the functionality available to extensions or plugins.
• Use whitelists, blacklists, or predefined action templates to control which tasks the system can perform.
• Set limits on the frequency and scope of actions initiated by the LLM to minimise cascading effects of malicious or unintended outputs.