Stockertim Dictionary · Artificial intelligence · Cybersecurity
Excessive agency
Also called over-permissioned AI, excessive AI permissions
In short
Excessive agency is giving an AI system more permissions, tools or autonomy than its task needs, which widens the harm a mistake or a misleading input could cause.
- Advanced
- Attack or threat
- Checked
In detail
An AI system can only do harm through what it is connected to. Excessive agency arises when that connection is broader than the job: an email-summarising assistant that can also send and delete mail, a reporting agent with write access to the finance system, or an agent allowed to act without approval on decisions that are hard to undo. The extra capability may never be needed, yet it is always available to be misused.
The risk comes from three directions. The model can make an honest mistake and take a wrong action. Hidden instructions in content it reads, a form of prompt injection, can steer it into a harmful one. And a flaw in the surrounding software can let a request through that should have been blocked. In each case the size of the damage depends on how much agency was granted.
The remedy is least privilege applied to AI: only the tools the task needs, each with the narrowest access, acting on behalf of the user with that user's rights, and approval for anything with significant or irreversible effect.
Worked example
Lagoon Logistics reviews an assistant that summarises the shared operations inbox.
- The review finds the assistant holds permission to read, send, forward and delete email.
- The task needs only reading, so send, forward and delete are removed.
- Drafted replies now go to a person's outbox for approval.
- A test with a harmless hidden instruction to forward mail now fails safely.
The assistant does its job with a fraction of the reach, and an injected instruction has nothing to act with.
In practice
Security reviews of AI features list every tool and permission, remove anything the task does not need, require approval for high-impact actions and recheck permissions whenever tools are added.
Watch out
- The risk is in what is permitted, not only in what the system normally does.
- A capable model does not reduce the need for least privilege; it can increase it.
Learn it properly
Official sources
- Top 10 for large language model applications · OWASP Foundation
- Guidelines for secure AI system development · National Cyber Security Centre
- Adversarial machine learning: a taxonomy and terminology of attacks and mitigations · National Institute of Standards and Technology
Facts last checked on 3 November 2026.
Education, not advice: the Stockertim Dictionary explains what terms mean. It never recommends buying or selling anything.