Quick answer
Excessive agency is when an AI system has more permissions, tools or autonomy than its task requires. It is the reason a successful prompt injection becomes a real incident: the more an AI agent can do, the more an attacker who controls it can do. It is listed in the OWASP Top 10 for LLM Applications.
On this page
How It Happens
Teams building agents tend to grant broad access so the agent can handle any request: full mailbox access instead of read-only, a database account that can write as well as read, an API token with administrator scope. Each extra permission is unused most of the time, until an attacker hijacks the agent through prompt injection or a poisoned tool.
Three Forms of Excess
- Excessive functionality: tools the agent does not need, such as delete or send functions.
- Excessive permissions: credentials with broader scope than the task.
- Excessive autonomy: sensitive actions taken without human confirmation.
How to Test for It
- List every tool, credential and permission each agent holds.
- Compare that list with what the agent's tasks actually require.
- Attempt to hijack the agent and use each unnecessary capability.
- Check which actions run without human approval.
- Report the blast radius: the worst realistic outcome of a compromised agent.
How to Reduce the Risk
Apply least privilege to agents exactly as you would to staff accounts, remove unused tools, scope credentials to the task, require confirmation for actions that send, delete, pay or change access, and log every tool call.
Regulatory and Framework Context
For Australian organisations, least privilege for AI agents extends the Essential Eight principle of restricting administrative privileges to a new class of identity.
Latest intel on this topic
- Gemini test proves once again that amoral AI agents will do almost anything if told ‘NEVER give up’The Register · 17 November 2026
- SpyCloud’s 2027 Forecast: Attackers Will Exploit Trusted Access, From Hijacked Sessions to AI AgentsHackerNoon · 17 November 2026
- Gemini AI agents repeatedly bypassed safeguards to attack real websiteBleepingComputer · 16 November 2026
- California sues OpenAI over AI agents’ summer hacking spreeCyberScoop · 16 November 2026
- California sues OpenAI over AI agents escaping sandboxes, hacking firmsBleepingComputer · 16 November 2026
Updated hourly from the daily briefing.
Common questions
Is an AI agent an identity?
In practice, yes. It holds credentials and acts on systems, so it should be governed like a service account: scoped, reviewed and monitored.
What is a human-in-the-loop control?
A requirement that a person confirms an AI-proposed action before it runs, used for sensitive actions such as payments, deletions and external messages.
Change log
- 13 November 2026First published.