AI Agent Risk Assessment
When an agent goes rogue, can you stop it?
Assess how prepared your organization is to recognize rogue behavior, enforce identity and business limits, and stop risky agent actions.
10 questions
Five short stages
Five readiness categories
Prioritized next steps
No account or email needed. Your answers stay in this browser.
What the assessment looks for
Agent governance happens at runtime.
01
Agents can be manipulated by attackers or go off task on their own.
02
Teams need to understand what agents actually do and intervene during execution.
03
Identity, access, security, and business policies must work together.
04
Controls should work across endpoints, SaaS, cloud, tools, and delegated agent workflows, without redesigning your AI architecture around security.
User request
“Clean up the old tables.”
Agent connects to the production database.
Allowed
Read access is permitted.
Allowed
Agent attempts to delete production tables.
Allowed access. Dangerous behavior.
Risky action
This is where control matters.
Can your policies stop the action before the damage is done?
How it works
Five short stages. Two questions each.
01
Stop risk
Spot an agent going off task and stop it before it causes harm.
02
Control access
Know whose account each agent uses and limit it to the access it needs.
03
Protect & trace
Keep sensitive data where it belongs and keep records you can show in an audit.
04
Test & discover
Test agents before launch and find AI tools running without approval.
05
Set limits
Check the tools agents connect to and decide what needs human approval.
Stop risk
Control access
Protect & trace
Test & discover
Set limits
Stage 1 of 5 · Stop risk
Spot rogue behavior
Would you know if an agent went off task?
Think about an agent making unexpected changes or acting beyond its assignment.
Stop risky actions
Can you stop an agent before it causes harm?
For example, before it deletes production data or shares customer information.
Know agent identity
Do you know whose account each agent is using?
This could be an employee account, a shared account, or a service account.
Limit access
Can you limit each agent to the access it needs?
For example, allow an agent to read records without letting it change or delete them.
Protect sensitive data
Can you prevent agents from sharing sensitive data where it doesn’t belong?
Think about customer records, credentials, and confidential company information.
Keep records for audits
Can you show what an agent did when someone asks for an audit?
Useful records connect the request, the account used, and the actions taken.
Test before launch
Do you test how safely an agent behaves before putting it to work?
Include attempts to trick it, such as prompt injection, as well as mistakes it could make on its own.
Find shadow AI
Can you find AI tools and agents being used without approval?
Consider employee devices, SaaS applications, and cloud environments.
Check connected tools
Do you check the security of tools and add-ons your agents use?
Examples include MCP servers, skills, plugins, and changes to those components.
Set business limits
Have you set limits on what agents can do without human approval?
For example, changing production systems, issuing refunds, or exporting large amounts of data.
Your answers, notes, owners, and dates are saved only in this browser on this device. The assessment doesn’t send them to Noma or anyone else.
Start overAI Agent Risk Assessment · Your results
Your results
Still to answer
Overall readiness
3
/ 5
Developing coverage. Higher ratings reflect more controls reported in place.
To verify
0
answers
Answers marked Not sure yet. They count as 1 until you confirm the control exists.
Your answers
Yes
0
Partially
0
Not yet
0
Not sure yet
0
Category scores
Five categories.Start with the shortest.
Each agent stands as tall as your 1–5 rating in that area. Lower agents are where to look first.
Overall readiness · 3 / 5
1
1
2
2
3
3
4
4
5
5
3
/5
Rogue & unintended actions
3
/5
Shadow AI
3
/5
AI supply chain
3
/5
Sensitive data exposure
3
/5
Agent manipulation
3
/5
Rogue & unintended actions
Recognize unexpected behavior, limit authority, stop risky actions, and keep records.
3 / 5
3
/5
Shadow AI
Find unapproved AI use and understand the accounts agents act through.
3 / 5
3
/5
AI supply chain
Check connected components and test agent behavior before use.
3 / 5
3
/5
Sensitive data exposure
Limit access and sharing, stop risky actions, and keep evidence of what happened.
3 / 5
3
/5
Agent manipulation
Recognize redirection, test attacks such as prompt injection, intervene, and review activity.
3 / 5
Q01
Question
Yes
What to do next
A few places to start
Start with areas to establish or verify, then build on controls already in progress.
Step 1
Establish the control
Action
Lead
How to start
How to start
Step 2
Establish the control
Action
Lead
How to start
How to start
Step 3
Establish the control
Action
Lead
How to start
How to start
Your answers
Responses and notes
Changing an answer updates your ratings and next steps. Owners and dates stay attached to their question.
Your 10 answers
5 stages · answers, evidence and notes
Show answers
Stage 1
Stop risk
Q01
Would you know if an agent went off task?
Not answered
Q02
Can you stop an agent before it causes harm?
Not answered
Evidence and notes
Stage 2
Control access
Q03
Do you know whose account each agent is using?
Not answered
Q04
Can you limit each agent to the access it needs?
Not answered
Evidence and notes
Stage 3
Protect & trace
Q05
Can you prevent agents from sharing sensitive data where it doesn’t belong?
Not answered
Q06
Can you show what an agent did when someone asks for an audit?
Not answered
Evidence and notes
Stage 4
Test & discover
Q07
Do you test how safely an agent behaves before putting it to work?
Not answered
Q08
Can you find AI tools and agents being used without approval?
Not answered
Evidence and notes
Stage 5
Set limits
Q09
Do you check the security of tools and add-ons your agents use?
Not answered
Q10
Have you set limits on what agents can do without human approval?
Not answered
Evidence and notes
Method
How your rating works
Yes = 5, Partially = 3, Not yet = 1. Not sure yet also counts as 1 and is flagged until you verify it. Each area averages its questions, and the overall rating weighs all five areas equally. Some questions count toward more than one area.
1 Getting started · 2 Building controls · 3 Developing coverage · 4 Broad coverage · 5 Strong reported coverage
This is a brief, self-reported readiness snapshot, not an audit, certification, or benchmark.A rating of 5 does not mean every control is in place.
Questions in each area
Rogue & unintended actions
Q01–Q04, Q06, Q07, Q10
Shadow AI
Q03, Q08
AI supply chain
Q07, Q09
Sensitive data exposure
Q02, Q04–Q06, Q10
Agent manipulation
Q01, Q02, Q06, Q07
Your answers, notes, owners and dates are saved only in this browser on this device.
Start overNext step
Found a gap?
Noma’s team can walk you through what you found and how to close it.