Visibility into AI use
Proxy, DNS and network logs build an AI inventory. Every object retains its source and the limits of observation.
CONTROLLED WORKFLOW
Connect → assess → explain → restrict
A pilot starts with agreed scope, sources and access permissions. Each step preserves results, limitations and history.
AS / SECURITY POSTURE
Illustrative architectureOne assessment perimeter for AI security
AS
AIScope
Select a diagram element to see its roleOne inventory connects systems, data, agents and tools. Assessments produce findings with sources, versions, confidence and coverage limits.
Define the scope
Select sources, allowed objects and permissions. Set the assessment profile, budget and schedule.
Build the inventory
Link AI systems, agents, tools and data to their owners, sources and current versions.
Assess and explain
Check configuration and content within scope. Preserve findings and evidence and identify objects that were not assessed.
Verify changes
Prepare restrictions, obtain approval and assess the result. Catalog or data changes require a fresh assessment.
RESULTA verifiable view of where AI is used, what is confirmed and which action comes next.
ASSESSMENT SURFACE
The whole AI system: from network signal to agent tool
Sources complement one another. Coverage depends on granted permissions and verified compatibility for each connection.
Unapproved AI use
Proxy, DNS and network logs help discover AI access outside approved routes. Suspected and confirmed bypasses remain distinct.
Data and knowledge bases
Assess RAG documents and chunks for sensitive data, injection indicators and access risks. Show which objects were read or unavailable.
Servers and tools
Assess MCP catalogs, transport, authentication, tool and resource descriptions. Metadata discovery does not automatically execute tools.
Agents, skills and permissions
Manifests, skills, dependencies and access relationships reveal agent risks in the context of available data and actions.
RESULTA pilot records specific sources and versions; a preset alone does not establish a verified integration.
VERIFIABLE FINDINGS
Every finding has evidence, a version and limits
Vulnerabilities, misconfigurations and semantic risks are assessed separately. AI explanations assist review; confirmation requires reproducible evidence.
Source and object
Findings retain the source, affected object, content or catalog version, assessment rule and observation time.
Confirmation level
A hypothesis, indicators and a reproducible assessment result are distinct. Unknowns remain visible.
Honest coverage
Unavailable, partially assessed and stale objects are explicit. No findings does not establish complete coverage.
Isolated active tests
Tool execution and dangerous action tests require separate authorization, test data and network, time and budget limits.
RESULTSecurity engineers receive reviewable evidence, system owners receive a clear next step, and auditors receive a decision history.
AISCOPE + AIGATE
From a finding to an approved restriction
AIScope observes and assesses. AIGate controls authorized LLM and MCP access and applies supported policies.
Prepare a change
A finding and catalog version support a policy candidate defining which tools, actions, arguments or resources to restrict.
Test the effect
Review the diff and simulate before application, using the capabilities of the installed AIGate version.
Obtain human approval
An authorized reviewer decides. A risk score or LLM output does not authorize policy application by itself.
Verify application
Check the active policy version and preserve history and rollback. Reassess risk when the inputs change.
RESULTDeployment in company infrastructure or private cloud; REST API and management MCP use shared permissions and services.