Skip to content

Roadmap

What is planned next. Delivered work is listed in Releases.

1. Sentinel Engram

Sentinel Engram is the long-term memory of Cyber Sentinel: a separate project that loads AI verdicts, DNS activity and Pi-hole blocklists into a Vertex Synapse CTI graph, with a Neo4j view for visual analysis. It reads Cyber Sentinel's database through a read-only role and an SSH tunnel and never writes to it.

Planned on the Cyber Sentinel side:

  • Ship the export in this repo — move db_engram_export.sql (role engram_reader, schema engram_export) and the engram-tunnel user into Cyber Sentinel's playbooks, so a dev deploy no longer removes them.
  • Graph context for the AI agent — n8n queries Synapse (user n8n-agent: read, ai.* tags only) to give the agent history the database does not have: blocklist membership, related domains on the same IP, first-seen dates.
  • Anomaly agent — a scheduled n8n agent compares daily device behaviour in the graph against a baseline and tags findings with ai.*. Device data must be anonymised before it leaves Cyber Sentinel.

2. Agentic Development

Goal: a system that checks itself, corrects its own mistakes and improves its own configuration, with a human approving changes instead of doing them. Each item runs as a separate n8n workflow; workflows hand work to each other through a status-driven queue table. Order below is the planned order.

  1. Health agent — daily check of the pipeline: queue depth, VirusTotal 429 errors, failed AI analyses, pg_cron jobs, aborted Tranco syncs, and host state from Prometheus (see section 6). Sends a short daily report and alerts when the pipeline stops.
  2. Verdict review and auto-unblock — weekly re-scan of verdicts scored 3–5 and of everything blocked in Pi-hole. Blocks get an expiry date; a domain that stays clean is unblocked, one that became malicious is raised.
  3. Analyst feedback — mark a verdict as false positive or true positive in the AI Config UI (or with the cs.fp tag in Engram). Labelled verdicts go into vector memory as confirmed examples.
  4. Prompt and threshold tuning agent — analyses verdict_audit and feedback: where the AI deviates from the rules and why, which rules produce false positives, how often the agent fails. Creates a new prompt version or setting change as an inactive draft, replays it against stored evidence from recent verdicts (no CTI API calls) and shows the difference. Activation stays manual in the AI Config UI.
  5. Retro-hunting — daily download of new ThreatFox IOCs, searched backwards in dns_queries: which device contacted a domain before it was known to be malicious. Needs no VirusTotal calls; can run in parallel with the items above.
  6. Agent-chosen enrichment — CTI sources become tools the agent picks within a budget instead of a fixed sequence. New sources: domain age (RDAP/WHOIS), certificate transparency (crt.sh), urlscan.io (already in Vault, unused).
  7. Engram tools — the Synapse graph as an agent tool and the device anomaly agent (see section 1). Optional at run time: the Engram host is up only a few hours a day, so anomaly analysis runs in batches.
  8. Agent-decided actions — blocking and alerting decided by the agent instead of fixed thresholds, with hard limits enforced by the workflow: never block trusted infrastructure, a daily block limit, every block with an expiry. Depends on item 2, so wrong decisions undo themselves.

Constraints for all items: VirusTotal free tier (500 requests/day), Gemini free tier limits, Raspberry Pi 5 with 8 GB RAM.

3. Device Attribution

  • user_devices table mapping internal IPs to devices and owners.
  • Per-device reports and alerts from n8n.
  • Identity signals (Vault audit log, SSH auth log) as a later step.

4. Hardening

  • TOTP for web UIs (n8n, Portainer, AI Config UI).
  • Nginx rate limiting and security headers (HSTS, X-Frame-Options, X-Content-Type-Options).
  • GeoIP blocking and port knocking for SSH.

5. File Integrity Monitoring

  • Hashes of critical files (sshd_config, playbooks, compose) stored in the database.
  • Scheduled comparison; an AI agent decides whether a change looks like administration or compromise and alerts.

6. Monitoring

  • Prometheus and node_exporter become the data source for the health agent instead of something to look at: disk usage, Pi temperature, SSD SMART (playbook 07), read by n8n through the Prometheus HTTP API.
  • Grafana will be removed. Its job — showing what happens in the pipeline — is taken over by the n8n agents: the health agent's daily report and alerts, plus operational views in the AI Config UI.