3-minute product tour Urgent incident email

AI for security operations
and security for AI systems

Use AI to triage alerts and gather evidence.
Evaluate the permissions and data flows of the chatbots and agents your organization operates.

AI gathers evidence.
Analysts review the result.

Automation levelAI roleHuman roleGenox configuration
L1 AssistanceSummarization and searchMake all decisionsAssisted workflow
L2 TriageAlert classification and prioritizationInvestigate and make decisionsTriage workflow
L3 InvestigationInvestigate and assemble the evidenceReview and approve verdictsDefault
L4 ResponseAutomate actions within the agreed policyDefine policy, handle exceptions and approve high-impact actionsConfigured by risk and asset

Unusual AI tool activity

Compare a chatbot’s tool calls with its normal workflow and check whether the requested action was authorized.

Ransomware indicators

Correlate unexpected file changes with evidence of backup interference before recommending containment for review.

Software supply-chain risk

Review a dependency change together with its publisher and provenance records.

Unexpected data access

Review an unusual download alongside access records and the account’s normal activity.

These scenarios are illustrative. See a sample verdict record for the evidence and review format.

Compare safeguards for a fictional assistant

A help-desk assistant can answer policy questions.
Explore how different controls affect an unexpected request for restricted information or a tool action.

Level 1 · Basic instructions

The assistant has instructions about what it may disclose. If it can access restricted information, a mistaken decision can still expose it.

Level 2 · Input screening

Screen incoming requests and retrieved content for known risk patterns. Staff can review ambiguous requests before the assistant proceeds.

Level 3 · Output screening

Check responses for restricted data before sending them. Review proposed tool actions separately so an approved message cannot authorize a tool call.

Level 4 · Restricted access and approval

Keep restricted data outside the assistant’s permissions. Treat retrieved material as data, and require independent approval for actions with material consequences.

This example uses predefined descriptions and does not connect to an AI model.

AI threats and relevant controls

ThreatWhat to assessIllustrative severityRelevant controls
Prompt injectionUntrusted content influences decisions or actions.HighTrust boundaries and behavior monitoring
Excessive agent permissionsAn agent can access more tools or data than its task requires.HighPermission inventory and least-privilege review
MCP supply chainConnected servers or packages have uncertain provenance or integrity.HighProvenance and signature verification, isolated assessment
Sensitive data exposureA response or retrieval step exposes restricted data.MediumRetrieval permissions and output screening
Model tampering and poisoningModel files or training data lack trustworthy provenance.MediumModel screening and an AI bill of materials
Shadow AITeams use AI systems outside the approved inventory.LowAI inventory and governance

Talk to Genox
about scope and timing.

Talk to us