Skip to content

armgit / AiSOC

Public Imported from GitHub
0

Open-source AI Security Operations Center: alert fusion, LLM-agent triage, MITRE ATT&CK investigation, and a replayable decision ledger for every agent step. Self-hostable, runs with no API keys, MIT licensed. Ships an MCP server for Claude, Cursor and Continue.

3.6 KB
Raw Blame

CVE Patch Policy

Status: approved by Alexandru Mateescu (2026-10-06). Implementation tracks the PR sequence at the bottom.

Problem

An open-source SIEM/HIDS's periodic vulnerability re-scan maps every finding to a new alert. A full scan of the four TMC agents produced 3,885 CVE alerts in two hours (2026-10-06 06:00–08:00 UTC); the queue stood at 3,902 new alerts, of which 1,576 were distinct CVEs. dedup_hash is keyed per event, so a re-scan of an unchanged host re-inserts every finding. The board becomes uncloseable and analysts stop trusting it.

Policy

Vulnerabilities are tracked findings, not incidents, with one exception.

  1. Servers get patched on the scheduled patch date. Always.
  2. A CVE with CVSS base score > 9, listed in CISA KEV, or flagged actively exploited is urgent: it must be patched out-of-band and it raises an alert immediately.
  3. Everything else is scheduled to the next patch date for the host's environment class. No alert at creation.
  4. A scheduled finding whose patch_due_date has passed while the scanner still reports it raises exactly one overdue alert.
  5. When the scanner stops reporting a finding, the row closes silently (remediated_at).

Patch calendar

Class Rule Next windows
TEST (hostname contains dev, test, uat) first Wednesday after Patch Tuesday 2026-10-14, 2026-11-11, 2026-12-09
PROD (default) test window + 7 days 2026-10-21, 2026-11-18, 2026-12-16

Patch Tuesday = second Tuesday of the month (the anchor; Linux vendors do not publish on a fixed date but distro kernel updates land in the same window).

Data model

  • asset_vulnerabilities (exists, currently Tenable-only) becomes the home for all scanner findings, keyed (tenant_id, asset_id, cve_id, source); re-polls move last_found only. Add environment_class derived from the hostname rule above, and patch_due_date + policy_status (scheduled|urgent|overdue|remediated).
  • The alerts table receives only: urgent creations (grouped per CVE, hosts as children), overdue promotions (one per CVE+host), and threat-intel re-promotions.
  • Open-source SIEM/HIDS vulnerability events stop creating per-finding alerts once the writer lands; the connector maps them into asset_vulnerabilities instead.

Day-one numbers (2026-10-06 snapshot, live DB)

  • 1,576 distinct open CVEs across 4 TMC agents (packages: linux-aws kernel).
  • 216 distinct CVEs with CVSS > 9 (500 CVE+host pairs) — urgent list.
  • 1,360 distinct CVEs in the 7–9 band — scheduled, due Oct 14 (test) / Oct 21 (prod).
  • 0 below 7.

The urgent list collapses to one action: patch + reboot the kernel on the four hosts.

Agent prompt (enforcement lives in the pipeline, not here)

Triage agents receive: "Vulnerability findings follow patch-window policy: severity ≤ 9 inside its patch window is scheduled maintenance, not an incident — never raise severity on it. Escalate only CVSS > 9, KEV, or findings overdue past patch_due_date."

PR sequence

  1. feat(alerts): POST /api/v1/alerts/bulk-close — ids or filter spec, max_count guard, backup table + audit entries (mirrors the manual Oct-2 bulk close that produced alerts_bulk_close_backup_20261002).
  2. feat(connectors): route open-source HIDS vulnerability findings to asset_vulnerabilities — per-event alert creation removed; urgent rule applied at write time.
  3. feat(api): patch_due_date + patch_windows config + overdue promotion job.
  4. feat(web): patch backlog view (CVE → affected hosts roll-up).
  5. chore(agents): patch-window policy paragraph in triage system prompt.