
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 level | AI role | Human role | Genox configuration |
|---|---|---|---|
| L1 Assistance | Summarization and search | Make all decisions | Assisted workflow |
| L2 Triage | Alert classification and prioritization | Investigate and make decisions | Triage workflow |
| L3 Investigation | Investigate and assemble the evidence | Review and approve verdicts | Default |
| L4 Response | Automate actions within the agreed policy | Define policy, handle exceptions and approve high-impact actions | Configured 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.
Assess AI before and after deployment
Pre-release AI security assessment
Review LLM applications, agents and model supply chains against the 20 assessment items. Document findings and recommended controls.
AI and LLM assessmentBehavior monitoring
Establish which tools an agent normally uses so analysts can investigate unexpected calls.
Behavioral AnalyzerAI management and governance
Review the AI inventory, risk assessments and approval procedures using Korea’s AI Basic Act and ISO/IEC 42001 as references.
Security consultingCompare 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
| Threat | What to assess | Illustrative severity | Relevant controls |
|---|---|---|---|
| Prompt injection | Untrusted content influences decisions or actions. | High | Trust boundaries and behavior monitoring |
| Excessive agent permissions | An agent can access more tools or data than its task requires. | High | Permission inventory and least-privilege review |
| MCP supply chain | Connected servers or packages have uncertain provenance or integrity. | High | Provenance and signature verification, isolated assessment |
| Sensitive data exposure | A response or retrieval step exposes restricted data. | Medium | Retrieval permissions and output screening |
| Model tampering and poisoning | Model files or training data lack trustworthy provenance. | Medium | Model screening and an AI bill of materials |
| Shadow AI | Teams use AI systems outside the approved inventory. | Low | AI inventory and governance |