Access & permissions
Repositories, data stores, credentials, tools and privilege boundaries.
AI agent security assessment
Secure the agent, not just the prompt.
AgentRiskLayer assesses AI agents that can use tools, APIs, MCP servers, data and autonomous workflows. We collect evidence, reproduce risk in an authorised boundary, guide remediation and retest the exact affected control.
Human-ledScope agreed firstNo secrets through this site
ARL Guardian · Public Preview
Map the authority surface before the assessment.
Guardian is a local, read-only capability discovery tool for AI-agent repositories. It maps reviewable evidence for MCP, outbound network access, filesystem mutation, process execution, credential references and approval boundaries without running the agent.
Guardian discovers. AgentRiskLayer assesses. Guardian does not decide severity, readiness, finding closure or deployment.
What we assess
Security around real agent authority.
Instruction attacks
Prompt injection, untrusted content, tool-result poisoning and instruction hierarchy.
Actions & approvals
High-impact actions, exact approvals, transaction boundaries and constrained targets.
Evidence & recovery
Auditability, containment, remediation evidence and exact retesting.
How it works
One bounded assessment workflow.
- 01Request & scope
We understand the system, lifecycle stage and authorised test boundary.
- 02Freeze & inspect
We bind evidence to the exact target revision and inspect the authority surface.
- 03Test & remediate
Controlled tests run only with explicit authorisation. Findings stay evidence-backed.
- 04Retest & report
The affected control is retested against the changed revision and reviewed by a human.
Decision boundary
The LLM is not the security authority.
AI can assist with explanation, organisation and drafting. It does not decide applicability, evidence validity, severity, finding closure, readiness or deployment.
Safe output ≠ safe action. A bounded PASS is not a universal security claim.
Request an assessment
Tell us what your agent can do.
We review the scope before any testing begins. Do not include passwords, API keys, access tokens, private keys or customer data.