Performing AI Driven Osint Correlation

Performing AI-Driven OSINT Correlation

When to Use

• You have collected raw OSINT data from multiple tools and sources but need to identify connections, contradictions, and patterns across them.
• You need to build a unified intelligence profile for a target entity (person, organization, or infrastructure) from fragmented data.
• Traditional manual correlation is too slow or error-prone for the volume of data collected.
• You want confidence-scored assessments of identity linkage across platforms rather than simple keyword matching.

Detection Gaps & Validation

• LLM hallucination and over-correlation: the model may invent linkages or assert connections the raw findings do not support. Never accept a correlation > 0.8 without independent verification against >=2 sources, and manually spot-check 10–20% of all linkages.
• Common-username false positives: generic handles (admin, test, info) and high-collision names inflate Sherlock matches into bogus entity profiles. Filter these before correlation, not after.
• Confidence-score inflation: treat the AI's 0.0–1.0 score as a hypothesis weighted by evidence type (exact username match high, temporal proximity low) - not ground truth. Require corroborating metadata for "medium" links.
• Stale/retracted OSINT: breach datasets and cached profiles persist after takedown. Verify timestamps are current and from reputable aggregators before including a finding.
• Source-bias gaps: Sherlock covers usernames, theHarvester emails/hosts, SpiderFoot infrastructure - a target active only on unmonitored platforms looks "clean." State coverage limits in the report.
• How to validate: confirm each linked account resolves independently, ensure no admin/test-type handle drove an entity merge, and verify the final profile excludes stale or contradicted data before assigning a risk level.

Prerequisites

• Python 3.10+ with requests, json, and csv libraries
• Sherlock installed (pip install sherlock-project)
• theHarvester installed (pip install theHarvester)
• SpiderFoot 4.0+ running on localhost:5001
• Access to an LLM API (OpenAI, Anthropic, or local model via Ollama)
• Optional: Maltego CE for graph visualization of correlation results
• Optional: API keys for Shodan, VirusTotal, HaveIBeenPwned, Hunter.io

Workflow

Legal & Ethical Requirements

• Obtain documented written authorization before any investigation
• Establish lawful basis for data processing (law enforcement, corporate policy, etc.)
• Define PII retention limits and data handling procedures
• Comply with local privacy regulations (GDPR, CCPA, etc.)

Phase 1 — Multi-Source OSINT Collection

0. Create the working directory for all OSINT outputs:

   mkdir -p /tmp/osint
   

1. Enumerate usernames across platforms with Sherlock:

   sherlock "targetusername" --output /tmp/osint/sherlock-results.txt --csv
   

2. Harvest emails, subdomains, and hosts with theHarvester:

   theHarvester -d targetdomain.com -b all -f /tmp/osint/harvester-results.json
   

3. Run a SpiderFoot passive scan via REST API:

   curl -s http://localhost:5001/api/scan/start \
     -d "scanname=target-recon&scantarget=targetdomain.com&usecase=passive" \
     | jq '.scanid'
   

4. Export SpiderFoot results when scan completes:

   SCAN_ID="<scanid_from_step_3>"
   curl -s "http://localhost:5001/api/scan/${SCAN_ID}/results?type=all" \
     -o /tmp/osint/spiderfoot-results.json
   

5. Query breach databases for email exposure (example with HIBP API):

   curl -s -H "hibp-api-key: ${HIBP_KEY}" \
     -H "User-Agent: OSINT-Correlation-Skill" \
     "https://haveibeenpwned.com/api/v3/breachedaccount/target@example.com" \
     -o /tmp/osint/breach-results.json
   

Phase 2 — Data Normalization

6. Normalize all collected data into a common schema. Create a unified JSON structure that tags each finding with its source, timestamp, and data type:

   cat > /tmp/osint/normalize.py << 'EOF'
   import json, csv, sys, os
   from datetime import datetime
   
   findings = []
   
   # Normalize Sherlock CSV results
   sherlock_path = "/tmp/osint/sherlock-results.txt"
   if os.path.exists(sherlock_path):
       with open(sherlock_path) as f:
           for row in csv.DictReader(f):
               findings.append({
                   "source": "sherlock",
                   "type": "social_profile",
                   "platform": row.get("name", ""),
                   "url": row.get("url_user", ""),
                   "username": row.get("username", ""),
                   "status": row.get("status", ""),
                   "collected_at": datetime.utcnow().isoformat()
               })
   
   # Normalize theHarvester JSON results
   harvester_path = "/tmp/osint/harvester-results.json"
   if os.path.exists(harvester_path):
       with open(harvester_path) as f:
           data = json.load(f)
           for email in data.get("emails", []):
               findings.append({
                   "source": "theHarvester",
                   "type": "email",
                   "value": email,
                   "collected_at": datetime.utcnow().isoformat()
               })
           for host in data.get("hosts", []):
               findings.append({
                   "source": "theHarvester",
                   "type": "hostname",
                   "value": host,
                   "collected_at": datetime.utcnow().isoformat()
               })
   
   # Normalize SpiderFoot results
   sf_path = "/tmp/osint/spiderfoot-results.json"
   if os.path.exists(sf_path):
       with open(sf_path) as f:
           for item in json.load(f):
               findings.append({
                   "source": "spiderfoot",
                   "type": item.get("type", "unknown"),
                   "value": item.get("data", ""),
                   "module": item.get("module", ""),
                   "collected_at": datetime.utcnow().isoformat()
               })
   
   with open("/tmp/osint/normalized-findings.json", "w") as f:
       json.dump(findings, f, indent=2)
   
   print(f"Normalized {len(findings)} findings from {len(set(f['source'] for f in findings))} sources")
   EOF
   python3 /tmp/osint/normalize.py
   

Phase 3 — AI-Driven Correlation

7. Send normalized findings to an LLM for cross-source correlation analysis:

   cat > /tmp/osint/correlate.py << 'PYEOF'
   import json, os
   from openai import OpenAI  # or anthropic, ollama, etc.
   
   client = OpenAI(api_key=os.environ["OPENAI_API_KEY"])
   
   with open("/tmp/osint/normalized-findings.json") as f:
       findings = json.load(f)
   
   correlation_prompt = f"""You are an OSINT analyst. Analyze these findings collected
   from multiple sources and produce a correlation report.
   
   For each identity or entity you detect:
   1. List all linked accounts/profiles with the evidence connecting them.
   2. Assign a confidence score (0.0-1.0) for each linkage based on:
      - Exact username match across platforms (high)
      - Similar usernames with shared metadata (medium)
      - Same email in breach data and registration (high)
      - Co-occurring infrastructure (IP, domain) (medium)
      - Temporal correlation of account creation dates (low-medium)
   3. Identify contradictions or potential false positives.
   4. Flag high-risk exposures (breached credentials, PII leaks, infrastructure overlaps).
   5. Produce a structured JSON report.
   
   Raw findings:
   {json.dumps(findings[:500], indent=2)}
   """
   
   response = client.chat.completions.create(
       model="gpt-4o",
       messages=[
           {"role": "system", "content": "You are an expert OSINT analyst specializing in identity correlation and link analysis."},
           {"role": "user", "content": correlation_prompt}
       ],
       temperature=0.1,
       response_format={"type": "json_object"}
   )
   
   report = json.loads(response.choices[0].message.content)
   
   with open("/tmp/osint/correlation-report.json", "w") as f:
       json.dump(report, f, indent=2)
   
   print(json.dumps(report, indent=2))
   PYEOF
   python3 /tmp/osint/correlate.py
   

8. Perform entity resolution — deduplicate and merge related identities:

   cat > /tmp/osint/resolve.py << 'PYEOF'
   import json
   
   with open("/tmp/osint/correlation-report.json") as f:
       report = json.load(f)
   
   # Extract entities and build a link graph
   entities = report.get("entities", [])
   print(f"Identified {len(entities)} distinct entities")
   for entity in entities:
       name = entity.get("identifier", "unknown")
       confidence = entity.get("confidence", 0)
       links = entity.get("linked_accounts", [])
       risk = entity.get("risk_level", "unknown")
       print(f"  [{confidence:.0%}] {name} — {len(links)} linked accounts — risk: {risk}")
   PYEOF
   python3 /tmp/osint/resolve.py
   

Phase 4 — Reporting and Visualization

9. Generate a final intelligence profile in Markdown:

   cat > /tmp/osint/report.py << 'PYEOF'
   import json
   from datetime import datetime
   
   with open("/tmp/osint/correlation-report.json") as f:
       report = json.load(f)
   
   md = f"# OSINT Correlation Report\n\n"
   md += f"**Generated:** {datetime.utcnow().isoformat()}Z\n\n"
   md += "## Entity Profiles\n\n"
   
   for entity in report.get("entities", []):
       eid = entity.get("identifier", "Unknown")
       conf = entity.get("confidence", 0)
       md += f"### {eid} (Confidence: {conf:.0%})\n\n"
       md += "| Source | Platform | Evidence |\n|--------|----------|----------|\n"
       for link in entity.get("linked_accounts", []):
           md += f"| {link.get('source','')} | {link.get('platform','')} | {link.get('evidence','')} |\n"
       md += f"\n**Risk Level:** {entity.get('risk_level', 'N/A')}\n\n"
       for flag in entity.get("flags", []):
           md += f"- ⚠️ {flag}\n"
       md += "\n"
   
   with open("/tmp/osint/intelligence-profile.md", "w") as f:
       f.write(md)
   
   print("Report written to /tmp/osint/intelligence-profile.md")
   PYEOF
   python3 /tmp/osint/report.py
   

10. Optional — Import correlation graph into Maltego for visualization:

    # Export entities as Maltego-compatible CSV for manual import
    cat > /tmp/osint/maltego_export.py << 'PYEOF'
    import json, csv
    
    with open("/tmp/osint/correlation-report.json") as f:
        report = json.load(f)
    
    with open("/tmp/osint/maltego-import.csv", "w", newline="") as f:
        writer = csv.writer(f)
        writer.writerow(["Entity Type", "Value", "Linked To", "Link Label", "Confidence"])
        for entity in report.get("entities", []):
            for link in entity.get("linked_accounts", []):
                writer.writerow([
                    link.get("type", "Alias"),
                    link.get("value", ""),
                    entity.get("identifier", ""),
                    link.get("evidence", ""),
                    link.get("confidence", "")
                ])
    
    print("Maltego CSV exported to /tmp/osint/maltego-import.csv")
    PYEOF
    python3 /tmp/osint/maltego_export.py
    

Key Concepts

Tools & Systems

Common Scenarios

• Threat Actor Attribution: Correlate a suspicious username found in a phishing campaign with social media profiles, domain registrations, and breach data to build an attribution profile.
• Attack Surface Mapping: Link discovered subdomains, emails, and employee social accounts to understand an organization's full external exposure.
• Insider Threat Investigation: Cross-reference an employee's known accounts with dark web marketplace activity and breach databases.
• Brand Impersonation Detection: Identify accounts across platforms mimicking a target brand by correlating registration patterns, naming conventions, and temporal signals.

Output Format

The final output is a structured JSON correlation report and a Markdown intelligence profile containing:

{
  "meta": {
    "target": "targetdomain.com",
    "sources_used": ["sherlock", "theHarvester", "spiderfoot", "hibp"],
    "total_findings": 247,
    "generated_at": "2025-01-15T14:30:00Z"
  },
  "entities": [
    {
      "identifier": "john.target",
      "confidence": 0.92,
      "linked_accounts": [
        {
          "source": "sherlock",
          "platform": "GitHub",
          "value": "john.target",
          "evidence": "Exact username match, bio references targetdomain.com",
          "confidence": 0.95
        }
      ],
      "risk_level": "high",
      "flags": [
        "Credentials exposed in 2 breaches (2022, 2023)",
        "Admin email for targetdomain.com found in public WHOIS"
      ]
    }
  ],
  "contradictions": [],
  "recommendations": []
}

Verification

• Confirm that each linked account has been independently verified against at least two sources before assigning confidence > 0.8.
• Cross-check AI-generated correlations manually for a random sample (10–20%) to validate accuracy.
• Verify that no false positives from common usernames (e.g., "admin", "test") inflated entity profiles.
• Ensure breach data timestamps are current and from reputable aggregators.
• Validate that the final report does not include stale or retracted OSINT data.