I'm a Security Team Assessing AI Risk

You are a CISO, security engineer, or compliance analyst. Developers on your team are using AI coding assistants, and you need visibility into what's running, what risks exist, and how to report on them.

Time estimate: 10 minutes per machine for initial discovery. 30 minutes for a full fleet assessment.

Prerequisites:

Node.js 18 or later
Access to developer workstations (directly or via remote execution)

Step 1: Discover Shadow AI

Run detection on a developer workstation to see what AI tools are active. This command is read-only and does not modify any files.

npx opena2a-cli detect

Expected output:

  Shadow AI Detection  v0.7.2

  Machine    dev-laptop-042 (darwin arm64)
  User       jsmith
  Directory  /home/jsmith/payments-api

  AI Agents (3 running)
  -----------------------------------------------
  claude-code          PID 41023   v1.0.12
  cursor               PID 38291   v0.45.6
  ollama               PID 52100   v0.3.14    local model

  MCP Servers (5 configured)
  -----------------------------------------------
  filesystem           claude     local
  postgres             claude     local
  github               claude     local
  slack                cursor     local
  jira                 cursor     local

  AI Config Files (3 found)
  -----------------------------------------------
  CLAUDE.md            project    governance rules
  .cursorrules         project    editor config
  .copilot/config.yml  project    copilot settings

  Governance Score     32 / 100

  Findings
  -----------------------------------------------
  - Project not registered (no AIM identity)
  - No SOUL.md behavioral governance file
  - 5 MCP servers ungoverned (no capability restrictions)
  - 3 credentials detected in source files
  - Config files unsigned
  - Local model (ollama) running without audit trail

  Run: opena2a protect    (fix all findings)

Key observations for security teams:

AI Agents shows every AI coding assistant running on the machine, including local models like Ollama that bypass corporate API gateways.
MCP Servers reveals what external services AI agents can access. A postgres MCP server means Claude can execute database queries. A slack MCP server means it can send messages.
Governance Score provides a single number (0-100) for executive reporting.

Step 2: Generate an Executive Report

Create an HTML report suitable for sharing with leadership or including in security reviews.

npx opena2a-cli detect --report

Expected output:

  Shadow AI Detection  v0.7.2

  Scanning dev-laptop-042...

  Report generated: shadow-ai-report.html

  Summary
  -----------------------------------------------
  AI Agents          3 running
  MCP Servers        5 configured
  Config Files       3 found
  Governance Score   32 / 100
  Critical Findings  2

  Opened shadow-ai-report.html in browser

The HTML report includes:

Executive summary with governance score
Full inventory of AI agents, MCP servers, and config files
Risk breakdown by category (credentials, governance, integrity, access)
Remediation steps ordered by severity
Machine metadata (hostname, OS, username, scan time) for audit trails

To save the report to a specific location without opening a browser:

npx opena2a-cli detect --report --no-open

Step 3: Export Asset Inventory

Export the discovery results as CSV for import into your CMDB, SIEM, or asset management system.

npx opena2a-cli detect --export-csv assets.csv

Expected output:

  Shadow AI Detection  v0.7.2

  Scanning dev-laptop-042...

  Exported 11 assets to assets.csv

  Breakdown
  -----------------------------------------------
  AI Agents          3
  MCP Servers        5
  Config Files       3

  Governance Score   32 / 100

The CSV file includes these columns:

hostname,username,scanDirectory,scanTimestamp,assetType,name,platform,scope,status,details
dev-laptop-042,jsmith,/home/jsmith/payments-api,2026-03-15T10:30:00Z,ai-agent,claude-code,,machine,running,PID 41023 v1.0.12
dev-laptop-042,jsmith,/home/jsmith/payments-api,2026-03-15T10:30:00Z,ai-agent,cursor,,machine,running,PID 38291 v0.45.6
dev-laptop-042,jsmith,/home/jsmith/payments-api,2026-03-15T10:30:00Z,ai-agent,ollama,,machine,running,PID 52100 v0.3.14 local model
dev-laptop-042,jsmith,/home/jsmith/payments-api,2026-03-15T10:30:00Z,mcp-server,filesystem,claude,local,configured,
dev-laptop-042,jsmith,/home/jsmith/payments-api,2026-03-15T10:30:00Z,mcp-server,postgres,claude,local,configured,
dev-laptop-042,jsmith,/home/jsmith/payments-api,2026-03-15T10:30:00Z,mcp-server,github,claude,local,configured,
dev-laptop-042,jsmith,/home/jsmith/payments-api,2026-03-15T10:30:00Z,mcp-server,slack,cursor,local,configured,
dev-laptop-042,jsmith,/home/jsmith/payments-api,2026-03-15T10:30:00Z,mcp-server,jira,cursor,local,configured,
dev-laptop-042,jsmith,/home/jsmith/payments-api,2026-03-15T10:30:00Z,config-file,CLAUDE.md,,project,found,governance rules
dev-laptop-042,jsmith,/home/jsmith/payments-api,2026-03-15T10:30:00Z,config-file,.cursorrules,,project,found,editor config
dev-laptop-042,jsmith,/home/jsmith/payments-api,2026-03-15T10:30:00Z,config-file,.copilot/config.yml,,project,found,copilot settings

Each row includes hostname, username, scanDirectory, and scanTimestamp so you can aggregate results from multiple machines into a single spreadsheet or database.

Step 4: Enrich with Community Trust Data

Cross-reference discovered MCP servers against the OpenA2A Trust Registry to see community trust scores, known vulnerabilities, and verification status.

npx opena2a-cli detect --registry

Expected output:

  Shadow AI Detection  v0.7.2

  Scanning dev-laptop-042...
  Enriching with Trust Registry data...

  MCP Servers (5 configured)
  -----------------------------------------------
  filesystem           claude     local      trust: 0.92   verified
  postgres             claude     local      trust: 0.87   verified
  github               claude     local      trust: 0.94   verified
  slack                cursor     local      trust: 0.78   unverified
  jira                 cursor     local      trust: --     not in registry

  Registry Findings
  -----------------------------------------------
  - slack: 2 known advisories (medium severity)
  - jira: not registered in Trust Registry (unknown provenance)

  Governance Score   32 / 100

The --registry flag queries the public Trust Registry API. Servers marked "verified" have had their identity cryptographically confirmed by the publisher. Servers not in the registry may be custom or internal tools -- they are not necessarily unsafe, but they lack community vetting.

Step 5: Full Security Review

Run a comprehensive 6-phase security assessment that combines credential scanning, config integrity, shadow AI detection, behavioral governance, advisory checks, and optional deep scanning.

npx opena2a-cli review

Expected output:

  OpenA2A Security Review  v0.7.2

  Project      payments-api v2.1.0
  Type         Node.js + MCP server
  Directory    /home/jsmith/payments-api

  Phase 1: Credential Scan
  -----------------------------------------------
  3 hardcoded keys found
    src/config.ts:12     ANTHROPIC_API_KEY    critical
    src/config.ts:15     OPENAI_API_KEY       critical
    src/db.ts:8          DATABASE_URL         high

  Phase 2: Config Integrity
  -----------------------------------------------
  3 config files unsigned
    CLAUDE.md            not signed
    .cursorrules         not signed
    .copilot/config.yml  not signed

  Phase 3: Shadow AI
  -----------------------------------------------
  3 agents running, 5 MCP servers configured
  Governance score: 32 / 100

  Phase 4: Behavioral Governance
  -----------------------------------------------
  No SOUL.md found
  0 / 54 ABGS controls addressed

  Phase 5: Advisory Check
  -----------------------------------------------
  1 advisory for installed MCP servers
    slack MCP: CVE-2026-1234 (medium)

  Phase 6: HMA Deep Scan (optional)
  -----------------------------------------------
  Skipped (run with --deep to enable)

  Summary
  -----------------------------------------------
  Critical     2    hardcoded API keys
  High         2    database URL exposed, unsigned configs
  Medium       3    no governance, MCP advisory, no identity
  Low          1    no runtime monitoring

  Security Score   30 / 100  -> 85 by running opena2a protect

  Report: security-review.html (opened in browser)

The HTML dashboard provides an interactive 6-tab view with drill-down into each finding. Each finding includes a description of the issue, a verification command to see it yourself, and a fix command to resolve it.

To generate the report without opening a browser:

npx opena2a-cli review --no-open --report security-review.html

For CI/CD integration, use JSON output:

npx opena2a-cli review --format json

Step 6: Aggregate Across Your Fleet

To assess AI risk across multiple developer workstations, run detection on each machine and combine the CSV exports.

Option A: Manual Collection

Run on each machine:

npx opena2a-cli detect --export-csv assets-$(hostname).csv

Then combine:

# On your analysis machine, after collecting all CSVs:
head -1 assets-dev-laptop-001.csv > fleet-inventory.csv
tail -n +2 -q assets-*.csv >> fleet-inventory.csv

Option B: Remote Execution

If you have SSH access to developer machines:

#!/bin/bash
# fleet-scan.sh
MACHINES="dev-laptop-001 dev-laptop-002 dev-laptop-003"
OUTPUT_DIR="./fleet-reports"
mkdir -p "$OUTPUT_DIR"

for machine in $MACHINES; do
  echo "Scanning $machine..."
  ssh "$machine" "npx opena2a-cli detect --export-csv /tmp/assets.csv" 2>/dev/null
  scp "$machine:/tmp/assets.csv" "$OUTPUT_DIR/assets-$machine.csv" 2>/dev/null
done

# Combine all CSVs
head -1 "$OUTPUT_DIR"/assets-*.csv | head -1 > "$OUTPUT_DIR/fleet-inventory.csv"
tail -n +2 -q "$OUTPUT_DIR"/assets-*.csv >> "$OUTPUT_DIR/fleet-inventory.csv"

echo "Fleet inventory: $OUTPUT_DIR/fleet-inventory.csv"
wc -l "$OUTPUT_DIR/fleet-inventory.csv"

Option C: JSON Output for SIEM Integration

Use JSON output and forward to your SIEM or log aggregation system:

npx opena2a-cli detect --format json | curl -X POST \
  -H "Content-Type: application/json" \
  -d @- \
  https://your-siem.example.com/api/v1/events

Interpreting Fleet Data

When reviewing aggregated data, look for:

Pattern	Risk	Action
Agents running without governance	Shadow AI	Deploy SOUL.md and AIM identity
MCP servers not in Trust Registry	Unknown provenance	Review and register or block
Local models (Ollama, LM Studio)	Data exfiltration bypass	Assess data handling policies
Low governance scores (<50)	Ungoverned AI usage	Prioritize for remediation
Credentials in source files	Secret exposure	Run `opena2a protect`

Compliance Mapping

OpenA2A detection and governance maps to common compliance frameworks:

Framework	Relevant Controls	OpenA2A Coverage
SOC 2	CC6.1 (Logical Access), CC7.1 (System Monitoring)	Shadow AI detection, runtime monitoring
ISO 27001	A.8.1 (Asset Management), A.12.4 (Logging)	Asset inventory export, event logging
NIST AI RMF	MAP 1.1 (AI inventory), GOVERN 1.1 (policies)	Detect command, SOUL.md governance
EU AI Act	Article 9 (Risk Management), Article 13 (Transparency)	Review command, ABGS controls

Next Steps

After completing the initial assessment:

Remediate critical findings by running opena2a protect on machines with hardcoded credentials
Establish governance baselines by deploying SOUL.md templates to all projects
Automate ongoing monitoring by adding detection to CI/CD pipelines (see CI/CD integration)
Track improvements by comparing governance scores over time

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

I'm a Security Team Assessing AI Risk

Step 1: Discover Shadow AI

Step 2: Generate an Executive Report

Step 3: Export Asset Inventory

Step 4: Enrich with Community Trust Data

Step 5: Full Security Review

Step 6: Aggregate Across Your Fleet

Option A: Manual Collection

Option B: Remote Execution

Option C: JSON Output for SIEM Integration

Interpreting Fleet Data

Compliance Mapping

Next Steps

Related Use Cases

FilesExpand file tree

security-team.md

Latest commit

History

security-team.md

File metadata and controls

I'm a Security Team Assessing AI Risk

Step 1: Discover Shadow AI

Step 2: Generate an Executive Report

Step 3: Export Asset Inventory

Step 4: Enrich with Community Trust Data

Step 5: Full Security Review

Step 6: Aggregate Across Your Fleet

Option A: Manual Collection

Option B: Remote Execution

Option C: JSON Output for SIEM Integration

Interpreting Fleet Data

Compliance Mapping

Next Steps

Related Use Cases