Authority
Can the agent only recommend, or can it write records, send messages, change permissions, deploy code or move money?
Evidence-based security review for tool-using AI
AI Agent Security Assessment
Run an AI agent security assessment that maps what the agent can access, what can influence it, what actions it can take and what evidence supports a deployment decision.
5–10 minutes for the first check
No card to start
£99 once for the full assessment package
What changes the risk
“AI agent” is not one security profile.
An assistant that drafts text is different from an agent with persistent memory, MCP tools, scheduled execution, customer data, financial authority or the ability to delegate work to other agents.
Influence
Can users, email, files, websites, tool output or stored memory inject instructions into the workflow?
Autonomy
Does it act only when asked, or can it run on schedules, react to events, replan, chain goals or continue without immediate approval?
Recovery
Can policy, memory, tooling, workflow and changed data be rolled back after a bad or manipulated action?
What the assessment covers
What an AI agent security assessment checks before deployment.
The review follows the same evidence chain used inside AgentRiskLayer. Capabilities can make controls relevant, but they do not become vulnerabilities by themselves.
PROFILE
Agent capabilities
Record autonomy, memory, tool discovery, delegation, trigger mode, learning or adaptation, evaluator authority and recovery scope.
ACCESS
Data and permissions
Review identities, tenant scope, customer data, secrets, files, databases, networks and administrative access.
TOOLS
MCP and action surface
Review tool trust, dynamic discovery, external services, write authority, action validation and downstream enforcement boundaries.
ATTACK
Prompt injection and misuse
Test authorised scenarios where untrusted input tries to influence tool calls, approvals, data access or protected actions.
CONTROL
Human approval and runtime policy
Check whether high-impact actions are bound to the exact action, target, parameters, value and validity period before execution.
PROVE
Evidence, remediation and retest
Connect observed controls, test evidence, findings, remediation ownership, exact retest and the resulting deployment decision.
Evidence model
A security claim is not the same as security evidence.
AgentRiskLayer keeps each stage separate so missing information is not silently converted into a vulnerability and a declared safeguard is not treated as verified control evidence.
- 1DECLARE
Declared controls and capabilities
The owner describes the exact agent version, its access, tools, autonomy, approvals and safeguards.
- 2OBSERVE
Observed controls and tests
Code, configuration, inspection output, authorised tests and runtime records are collected separately from declarations.
- 3FIND
Findings only when supported
A finding requires an observed or reproducible failure. Unknown, missing or inconclusive information remains an information gap.
- 4DECIDE
Fix, retest and decide
Remediation is recorded, the exact failure is retested and current evidence supports a proceed, hold or do-not-deploy decision.
Example assessment question
An MCP-enabled support agent can issue refunds. What proves the action is controlled?
A useful answer needs more than “human approval is enabled.” The assessment checks where approval is enforced, what exact action and value are bound to it, whether it can be replayed, what happens when parameters change and what evidence records the decision.
Assessment rule
PROVE IT
Capability → applicable control → test → evidence → finding only if failure is demonstrated.
- Unknown information?
- Context required
- Failed test?
- Finding path
- Fix recorded?
- Exact retest
Start with one agent
Free qualification first. Evidence workflows when you need them.
The free check identifies what needs review. The £99 AI Agent Security Assessment unlocks the full report and customer-operated inspection, controlled-testing, remediation and retest workflows. It reports only work actually completed.
Free agent check
£0Describe the agent, its access and safeguards and get an initial risk view.
Start freeAI Agent Security Assessment
£99 onceFull report and PDF plus evidence, controlled-testing, remediation and exact-retest workflows for one agent.
Start with the free checkRuntime protection
From £29 /monthOngoing projects, runtime decisions, retention and team workflows.
Compare plansQuestions buyers and builders ask
AI agent security assessment FAQ
Is this a certification?
No. AgentRiskLayer Security Assessment is a proprietary assessment against the AgentRiskLayer Control Profile. It is not an accredited certification or a guarantee that a system is risk-free.
Does an unknown capability become a finding?
No. Unknown or unconfirmed information remains context. A finding requires an observed or reproducible control failure.
Can the assessment cover MCP agents?
Yes. The assessment can review MCP and other tool-enabled agents, including tool trust, permissions, external dependencies, action validation, prompt injection exposure and approval boundaries.
What happens after a finding is fixed?
The remediation is recorded against the changed system version and the exact original failure is retested before closure.
Do I need to expose secrets?
No. The assessment is designed around privacy-safe descriptions and evidence. Do not enter credentials or secrets into the assessment fields.
Where can I see the methodology?
Read the assessment methodology, Trust Centre and sample report before starting.
Assess one real agent
Start the AI agent security assessment
Find out what it can do and what evidence you still need.
Start free. Unknowns remain unknown. Findings require evidence.