| Infrastructure Category | Severity | Feb 23 | Feb 24 | Delta | Change | Trend |
|---|---|---|---|---|---|---|
| Exposed Ollama Instances | CRITICAL | 17,536 | 17,533 | −3 | −0.02% |
|
| Unsecured Jupyter Notebooks | HIGH | 111 | 5,415 | +5,304 | +4,779% |
|
| Open WebUI Interfaces | CRITICAL | 744 | 821 | +77 | +10.3% |
|
| Open n8n Automation Servers | HIGH | 41 | 40 | −1 | −2.4% |
|
| LangChain API Endpoints | MEDIUM | 7 | 1 | −6 | −85.7% |
|
| AutoGPT Instances | MEDIUM | 4 | 0 | −4 | −100% |
|
| TOTAL EXPOSURE | 18,443 | 23,810 | +5,367 | +29.1% |
Jupyter Notebook Spike — 111 → 5,415
The 48-hour jump from 111 to 5,415 exposed Jupyter instances is the statistical outlier in this scan set. Three explanations are worth evaluating before this number is published:
1. Query refinement artifact. The Feb 23 query may have used a narrower filter (e.g. strict product match) versus a broader port-only query on Feb 24. If the Shodan query string changed between scans, the delta reflects methodology, not reality. Worth validating by re-running the exact Feb 23 query.
2. Shodan re-index event. Shodan periodically re-crawls IP ranges it hasn't seen recently. A scheduled re-index of certain cloud regions (AWS, GCP, Azure edu tiers where Jupyter is common) could surface thousands of hosts simultaneously — not newly exposed, just newly catalogued.
3. Genuine overnight spike. Less likely but possible: a widely-shared tutorial, conference workshop, or auto-deploy tool pushed a large volume of Jupyter environments into production without hardening. This happens around major AI conferences and hackathons.
The remaining categories are stable. Ollama is effectively flat (−3), confirming it as a persistent structural risk rather than a trending one. Open WebUI grew +10.3%, a steady climb consistent with increased self-hosted LLM adoption. LangChain and AutoGPT declining to near-zero likely reflects Shodan fingerprint drift as those projects evolve their exposed interfaces.