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.
AiSOC Roadmap
📌 Active planning has moved (2026-05-12)
This file is the historical record of major-version deliverables (v4 → v8 planned). Day-to-day planning, prioritization, and contributor-facing issue intake now live in the community-feedback-driven Now / Next / Later docs:
docs/community-feedback/2026-05-12/AiSOC_ROADMAP.md— strategic narrativedocs/community-feedback/2026-05-12/AiSOC_Community_Feedback_Synthesis.md— themes (F001–Fxxx)docs/community-feedback/2026-05-12/AiSOC_Proposed_Issues.md— 23 implementation ticketsItems in v7.1+ sections below that overlap with the new docs are flagged inline with
→ Now/Next/Later: [ID]. Where the new docs supersede a deferred item entirely, the entry here is left intact for traceability but reasoned about against the newer plan.
This document captures the planned direction for AiSOC across major versions. All v4 deliverables and items deferred beyond v4 are listed here.
World-Class Hardening Program (2026-07, in flight)
A proof-first, security-first program to make every README claim gate-backed, close four existential agent-security holes, and build the ingest-time-graph + multi-model-router moat. Executed one phase per PR with a mandatory CI gate each. Committed status is the checklist below (the six lettered deferrals — 3.5+, 5b, 7b+, 9b, 10b, 11b — are scoped in docs/audit/DEFERRED_SUBPHASES.md. They were previously said to be "tracked in docs/audit/PROGRESS.md", which is gitignored and was never committed, so six named commitments had no scope anywhere a contributor could read). Baseline audit: docs/audit/REALITY_REPORT.md and docs/audit/CLAIM_TO_GATE_MATRIX.md.
- Phase 0 — Reality audit (claim-to-gate matrix, ranked overclaims/untested-paths/circular-gates)
- Phase 1 — Four existential holes (prompt injection, memory poisoning, cross-store tenant isolation, data-exfiltration/redaction, cost DoS, vault)
- Phase 2 — Supply chain + truth (security.yml scanners, claim-gate ratchet, hard-fail insecure prod defaults, TRADEMARK, verifying-releases, license fix; continuation landed: per-image cosign signatures + CycloneDX SBOM attestations + SLSA provenance, all actions SHA-pinned)
- Phase 3 — Integration / E2E / chaos / DR (real-container spine test, backup-restore, chaos, upgrade, cross-store isolation live-replay; heavy-demo-stack Playwright E2E + demo-timing gate tracked as non-blocking 3.5+)
- Phase 4 — Real evals + detection content truth table (third-party-labeled corpus, hallucination/calibration/abstention, model matrix) — 4a/4b/4c landed: de-circularised DAC candidate-rule gate; honest executable-vs-imported truth table; hallucination, abstention, calibration and containment metrics in
packages/aisoc-benchmarkwith a documented adapter so a third-party agent can be graded on the same corpus; confidence calibration gated intest_confidence_calibration.py; and a model matrix (scripts/run_model_matrix.py) wired into the weekly wet eval, which grades the same corpus across several models by re-invoking the existing evaluator rather than defining a second notion of accuracy. Remains open on one thing only, and it is not code: the live-agent numbers need a funded provider key. Without one the matrix reports not measured per model rather than emitting zeros, because a zero is a measurement and "we did not run this" is not. - Phase 5 — Data spine correctness (versioned event-schema registry + dead-letter queue + source-event lineage in the fusion consumer; idempotency via AlertSink dedup + event-time watermarking; backfill/replay-from-offset tracked as 5b)
- Phase 6 — Performance + cost (fusion hot-path throughput harness with a generous regression-floor gate; deterministic storage $/TB cost model + drift gate; storage-consolidation ADR-0005)
- Phase 7 — Ingest-time graph + multi-model router — graph-at-ingest already shipped (v8 T1.1,
services/ingest/internal/graph/); 7a landed: unified deterministic→ML→LLM router with tier attribution + theAISOC_DETERMINISTICdeterminism contract (services/agents/app/routing/model_router.py, gated). 7b+ (posture collection, effective-permissions snapshot loader, bi-temporal valid_from/valid_to, fusion-time ContextBundle) scoped indocs/audit/DEFERRED_SUBPHASES.md, which records that all five effective-permissions resolvers already ship and that what remains is one thing: no connector answers__posture_snapshot__ - Phase 8 — LLMOps (version+hash-pinned prompt registry with a CI drift gate; model pins + deterministic-terminated provider-fallback chains; content-addressed response cache; fail-closed structured-output validation)
- Phase 9 — Autonomy safety (dry-run-by-default policy, honest bounded rollback-capability contract replacing silent
return True, mandatory post-action verification for unattended containment, break-glass HIGH-blast flag, autonomy scorecard — all gated at the policy layer; live-router wiring + durable approval-SLA timer table tracked as 9b) - Phase 10 — Connector + content quality (runtime-contract conformance suite across all 69 connectors registered when the phase landed; the generated
conformance-matrix.mdnow reads 84 / 84 — gates the "live Test connection" capability that had NO gate + secret-field-marking; the matrix is published with a drift gate on it; detection lifecycle already gated by Phase 4a DAC + Phase 4b truth table. Live-vendor sandbox smoke + rate-limit/checkpoint durability tracked as 10b) - Phase 11 — API / SDK / release engineering (pure-Python OpenAPI breaking-change detector
scripts/openapi_diff.py+openapi-breaking.ymlgate: PR spec vs base fails on removed endpoint/schema/field, type change, tightened request, or dropped enum value — closes the OpenAPI NO GATE row; per-language SDK generated-client contract-drift tracked as 11b) - Phase 12 — Observability + governance (per-service SLOs in
docs/operations/slos.yamlwith a coverage gate;docs/operations/observability.mddocumenting the four golden signals + single OTel trace across the spine;GOVERNANCE.md+MAINTAINERS.md+ DCO sign-off inCONTRIBUTING.md; governance-completeness gate —governance.yml)
Program status: all 13 hardening phases (0–12) are landed. Building on them, the Fully-Operational AI-SOC roadmap (Phases A1–E1) is now complete — it wired the three end-to-end paths the reality audit found unwired and pushed the platform to competitive parity + beyond:
- Phase A (SIEM foundation): A1 ClickHouse lake writer · A2 live detection-evaluation worker (2538 executable rules on the stream) · A3 default-on one-command deploy (connectors + graph on) · A4 UEBA behavioral-model fusion.
- Phase A (SIEM foundation): A1 ClickHouse lake writer · A2 live detection-evaluation worker (2538 executable rules on the stream) · A3 default-on one-command deploy (connectors + graph on) · A4 UEBA behavioral-model fusion.
- Phase B (SOAR foundation): B1 auto-triage worker (copilot default) · B2 credential resolver +
decide()-governed live dispatch + 10 vendor adapters · B3 real rollback + post-action verification + durable approval-SLA timers · B4 Business Context Rules in the hot path. - Phase C (parity differentiators): C1 Advanced Data Explorer · C2 Effective-Permissions live posture loader · C3 autopilot/copilot posture scorecard · C4 fuse-time attack-chain auto-grouping.
- Phase D (breadth): D1 eight connectors (QRadar/Exabeam/Securonix/Devo/Netskope/Windows-Sysmon/Zeek-Suricata/syslog-CEF) · D2 AI/LLM-usage audit connector +
llm-*detections + hot/cold lake tiering · D3 live-vendor mock-server smoke. - Phase E (prove it): E1 CI-gated benchmark scoreboard tied to a deterministic live-agent MITRE-accuracy run.
The claim-to-gate matrix stands at 350 rows — 345 GATED / 5 PARTIAL / 0 NO GATE. Five rows are PARTIAL, each naming what closes it: "Cloud-native collection" until a capture recorded from a live Event Hub replaces the fixture synthesised from the Avro specification; "the retention window a tenant chooses is the one the lake honours" until a live ClickHouse test watches a row survive past the old 90-day boundary; "the AI estate is discovered on the endpoint" until a containerised osquery runs the shipped pack end to end; and "a reported phishing message can be purged across its recipients" until a maintainer-supplied Microsoft 365 or Google Workspace tenant lets the documented request shapes be exercised for real; and "every automatic closure of a production alert is checked against the current policy, grant and stop controls" until the live suite that proves it against real Postgres has been observed to pass on main rather than only in the pull request that added it. "SSO completes a sign-in" closed the way the programme requires — by building the gate it named: sso-live.yml stands up a containerised Keycloak and both legs have been observed to pass. Every other product claim is backed by a failing test, none is a named deferral either, and the ratchet (MAX_NO_GATE=0) forbids any regression. Every PARTIAL row was closed by building the gate it named, never by relabelling. Count the table rows with python3 scripts/check_claim_gate_matrix.py rather than trusting a figure quoted in prose — this line has gone stale before. The six lettered deferrals are scoped in docs/audit/DEFERRED_SUBPHASES.md.
v4.0 — Shipped
AI multi-agent investigator
- Orchestrator (LangGraph state machine) in
services/agents/app/investigator/ - ReconAgent, ForensicAgent, ResponderAgent (dry-run with analyst approval)
- ReportWriterAgent — streaming markdown + branded PDF
- Investigation & Report tabs in Case Workspace UI
- Eval harness: 20 synthetic incidents, ≥80% MITRE-tactic accuracy CI gate
Visual SOAR studio
- React Flow playbook editor over nine of the step types the engine implements (enrich, investigate, notify, block_ip, isolate_host, create_ticket, close_case, http, condition). The engine accepts 22 and runs 21; widening the palette to the rest is outstanding. Loop, Parallel, Wait and Human Approval were listed here and exist in neither the palette nor the engine.
- Sequential playbook engine with conditions, branching, retries and
cycle detection. Not a DAG: there is no
depends_onand no parallel execution, and there are no idempotency keys. - Step-level risk grading, by dispatch rather than by a schema field.
A step naming a response verb is dispatched to the action registry and
graded against that verb's capability contract — impact, reversibility,
approval requirement, verification probe — plus the tenant's autonomy
tier and the finding's confidence. Per step, so authorising a playbook
does not authorise what its steps contain. The old
blast_radiusstep field promised this and was declared by a schema the engine could not read; it has been removed rather than left as a promise. -
schemas/playbook.schema.json(JSON Schema draft-07) for portability and CI linting, held to the engine in both directions byscripts/check_playbook_schema_parity.py - Detection-as-Code:
detections/directory with Sigma + AiSOC YAML, GitHub Action deploy-on-merge - 12 starter playbook templates
- Community playbook marketplace (static index v4.0; publishing flow v4.1)
Plugin platform, public API, SDKs, docs
- Plugin SDK in Python (
packages/plugin-sdk-py/) and Go (packages/plugin-sdk-go/) -
plugin.yamlmanifest spec (connector | enricher | responder | detection | widget) - Plugin loader with OCI image support (
oras pull) in api/actions/enrichment/connectors - Public REST API v1 at
/api/v1, OpenAPI 3.1 atdocs/openapi.yaml - GraphQL gateway (Strawberry) proxying REST
- Scoped API tokens (
cases:read,playbooks:run,plugins:install) - Auto-generated client SDKs:
@aisoc/sdk(TypeScript),aisoc-sdk(Python/PyPI),github.com/beenuar/AiSOC/packages/sdk-go - Docusaurus docs site at
docs/site/, deployed to GitHub Pages - Demo Lab:
pnpm aisoc:labone-command full-stack + Conti-style ransomware scenario - 4 reference plugins: Okta connector, YARA enricher, Slack quarantine responder, MTTR sparkline widget
Cross-cutting
- OpenTelemetry traces: agents → actions → api → realtime (Jaeger/Tempo)
- API token scopes (foundation for SSO)
-
docs/upgrade/MIGRATION.mdfor v3 → v4 upgrade path
v4.1 — Shipped
- Plugin publishing flow (signed community submissions, Ed25519 verification, review endpoints)
- Plugin marketplace UI v2 (ratings, install counts, verified badges, category filter, sort)
- Detection catalog: browse and install community Sigma rules via UI
- Playbook community submissions and curation
-
aisoc-cli— developer CLI for scaffold, validate, publish plugins and detections
v5.0 — Shipped
Identity & Access
- SAML 2.0 + OIDC authentication (Okta, Azure AD, Google Workspace)
- Multi-tenant row-level security (Postgres RLS + SQLAlchemy middleware)
- Granular RBAC with data-class and tenant scopes (
require_permission()dependency) - Full analyst audit log (append-only
audit_logtable + middleware + UI)
Compliance
- Compliance evidence mapping for 24 controls across 5 frameworks (SOC2, PCI-DSS, HIPAA, ISO27001, NIST-CSF), with PDF export. Not a SOC 2 Type II evidence dashboard: the console pages for it call routes that do not exist. Restored by parity 1.3
- ISO 27001 control mapping
- NIST CSF / NIST 800-53 control coverage heatmap
- PCI-DSS and HIPAA control mappings. DORA is not mapped by any code path
- MTTD / MTTR / MTTC SLA tracking per tenant
High Availability & Operations
- HA Helm chart with PodDisruptionBudgets and HorizontalPodAutoscalers
- Backup / restore CLI (
scripts/backup.sh,scripts/restore.sh) - Multi-region active-active topology guide (
docs/operations/multi-region.md) - Operator runbook generation from OTel traces (
scripts/generate_runbook.py)
v5.1 — Shipped
UEBA
- Per-user, per-host, per-service behavioral baselines (Welford's algorithm)
- Anomaly risk scores feeding the fusion engine (z-score composite scoring)
- Peer-group analysis and deviation scoring
- Kafka integration: consumes
security.events, publishesueba.anomalies
Deception / Honeytokens
- Token generation (AWS keys, URLs, DNS, file, DB credentials, custom types)
- First-touch alerting via HMAC-SHA256-signed webhooks
- Honeytoken lifecycle management UI (create, revoke, delete, trigger history)
Purple-Team / Continuous Validation
- Atomic Red Team YAML loader and test sync API
- Caldera adversary emulation REST client integration
- ATT&CK coverage heatmap by tactic/technique with detection tracking
- Tabletop incident simulator with findings management UI
v6.0 — Shipped (2026-05-06)
Wave 3 — Operational Maturity
- MSSP / parent-tenant console — onboard child tenants, delegate cross-tenant actions, view rollup metrics
- Asset inventory + vuln-to-alert correlation — asset CRUD, vulnerability findings, blast-radius context
- Insider threat module — user risk profiles, behavioural indicators, peer-group deviation scoring
- L0–L4 auto-remediation maturity tiers — per-tenant autonomy gate with audit log and per-action whitelist
Wave 4 — Advanced Capabilities
- Internal threat intelligence — IOC harvesting, threat actor profiles, STIX/TAXII feed subscriptions
- Cloud security posture management (CSPM/KSPM) — posture findings, drift tracking, suppress/resolve workflows
- Identity-centric correlation graph — identity node/edge graph, alert-to-identity linking, attack-path queries
- Auto-generated board reports — report templates, scheduled PDF/HTML artefacts, email/webhook delivery
Platform
- Dashboard metrics API — aggregated KPI endpoint powering frontend dashboard tiles
- Tailscale connector — audit log and policy-change events with cursor-based pagination
- AWS GuardDuty credential-exfiltration Sigma detection rule
v6.1 — Shipped (2026-05-07) — v1.5 market-driven feature expansion
A review of G2, Gartner Peer Insights, and customer feedback on AI SOC / SIEM / SOAR platforms drove this release.
New autonomous agents (services/agents/app/agents/)
- Master autonomous triage agent (
auto_triage_agent.py) — classifies each alert astrue_positive/false_positive/benignwith confidence - Phishing triage sub-agent (
phishing_agent.py) - Identity reasoning sub-agent (
identity_agent.py) - Cloud reasoning sub-agent (
cloud_agent.py) - Insider-threat reasoning sub-agent (
insider_threat_agent.py) - All five exposed via
POST /api/v1/agents/triage
New console pages (apps/web/src/components/)
-
/investigate— redirects to/hunt; the multi-turn copilot isCopilotDockand/copilot -
/coverage-advisor— ATT&CK techniques your rules reference, ranked by enabled coverage (not by adversary prevalence) -
/shifts— analyst shift-handoff dashboard -
/easm— External Attack Surface Management -
/mssp— MSSP executive dashboard -
/noise-tuning— per-rule false-positive rate and one-click tuning -
/analytics/team— analyst leaderboard, MTTR per analyst, dispositions accuracy
New API surfaces (services/api/app/api/v1/endpoints/)
-
shifts.py— shift-handoff CRUD -
stix_taxii.py— STIX 2.1 / TAXII 2.1 publishing -
compliance.py, automated compliance evidence for 24 controls across SOC 2, ISO 27001, NIST CSF, PCI-DSS and HIPAA. DORA is not among them -
deployment.py— deployment / air-gap toggles
New connectors (16 → 26)
- SentinelOne (
sentinelone.py) - Cortex XDR (
cortex_xdr.py) - Wiz (
wiz.py) - Snyk (
snyk.py) - Zscaler (
zscaler.py) - Proofpoint (
proofpoint.py) - ServiceNow (
servicenow.py) - Jira (
jira.py) - 1Password (
1password.py) - Duo Security (
duo_security.py)
Other
- AI-generated incident reports — one-click "Export Report" generates PDF from the Investigation Ledger
- Air-gap deployment configuration — per-tenant toggles disable external feeds
v7.0 — Shipped ✅ (2026-05-10)
All items below were shipped as part of the v1.0 buyer-value plan. Implemented and reviewed by Beenu Arora [email protected].
- axe-core CI gate over the landing and chrome components plus three of the five operator views parity 4.7 names: the alerts queue and list, the investigation rail, and settings. Still not a full WCAG AA pass: the case workspace is not covered, and axe catches a subset of WCAG rather than all of it. Each view asserts it rendered real markup before axe runs, because axe passes on an empty div
- Light theme persisted in user profile (
ThemeProvider.tsx+PATCH /api/v1/users/me/preferences) - Saved views and custom drag-drop dashboard widgets per analyst (
saved_views.py+DashboardView.tsx) - AI-generated weekly executive digest — auto-emailed PDF (
digest_pdf.py+weekly_digest_task.py) - Slack native bot for alert triage without opening the UI (
services/slack-bot/— 61 tests) - Threat actor attribution engine v0 (
services/threatintel/app/actors/attribution.py) - Air-gap / Ollama local-LLM mode (
infra/compose/docker-compose.airgap.yml+apps/docs/docs/operations/air-gapped.md) - BYOK per-tenant LLM credentials UI + API (
llm_credentials.py+SettingsView.tsx) - MSSP console — per-child-tenant KPI aggregation, SLA posture, parent_tenant_id hierarchy
- Team analytics view — analyst MTTR, leaderboard, shift workload (
TeamAnalyticsView.tsx) - Case auto-summary + PDF export (
case_summary.py+case_summary_html.py) - Investigation timeline (replayable) (
InvestigationTimeline.tsx) - Playbook gallery with 12 curated packs + GitHub PR integration for detection proposals
- Mobile responder console — decide an approval from a phone
(shipped in v9.0. The "not started" note this line used to carry was
true about React Native and misleading about the product: the responder
console already existed as a PWA — nine routes under
apps/web/src/app/(responder)/, a service worker, an IndexedDB offline approval queue, Web Push with VAPID, and passkeys.apps/mobileis a distribution channel on top of that, and it exists for one reason worth the build: iOS Web Push requires an installed PWA and has been unreliable even then, which for "approve a containment from your phone" is the same as the feature not existing. Itssrc/libunit tests and type-check run in CI; no device build, simulator run or store submission has been performed, and APNs/FCM credentials are an account action rather than an engineering one — seeapps/mobile/README.md.) - Plugin publishing marketplace v3 (commercial plugins, revenue sharing) (open and unscheduled, and never in fact scheduled against a particular release. The "deferred past v8.0" this line used to carry dated the label rather than the decision, and survived five majors. Revenue sharing is a commercial decision rather than an engineering one, and the free packaging path is itself still blocked on registry credentials — see v8.1.)
v7.0.x — Endpoint telemetry wave + hardening (2026-05-10)
Six-PR feature wave that closes #44 ("osctrl connector for fleet-wide osquery telemetry") and significantly extends osquery coverage end to end. All six PRs were implemented sequentially as part of the v7.0 release window and then patched through 7.0.1 → 7.0.3.
Endpoint telemetry — osquery feature wave (PR1–PR6)
- PR1 — osctrl + FleetDM connectors (
services/connectors/app/connectors/osctrl.py,fleetdm.py). Schema-driven setup, liveTest connectionround-trip, secrets encrypted viaCredentialVault, polling on per-instance schedule, plus marketplace manifests atplugins/osctrl/plugin.yamlandplugins/fleetdm/plugin.yaml. - PR2 — Native osquery detection schema migration — 16 osquery rules
(
detections/endpoint/osquery-*.yaml, IDsdet-endpoint-281..296) migrated from_quarantine/to the native schema, with positive/negative test fixtures (detections/fixtures/osquery_*.json) gated by the Detection Validation workflow. - PR3 — Live-query playbook step (
services/actions/app/clients/osctrl_client.py,fleetdm_client.py,osquery_allowlist.py,services/agents/app/playbook/steps/osquery_live_query.py). Allowlisted distributed queries pushed to single hosts or fleet-wide via osctrl/FleetDM with HMAC-signed ChatOps approval. - PR4 —
aisoc-osquery-tlsFastAPI service +aisoc-directagent connector (services/osquery-tls/,services/connectors/app/connectors/aisoc_direct.py). First-party self-hosted osquery TLS plugin, FleetDM-compatible config/log endpoints, direct-from-agent ingest path that bypasses third-party SaaS. - PR5 — Osquery packs + FIM endpoint + FIM dashboard
(
services/osquery-tls/app/api/v1/endpoints/fim.py,apps/web/src/components/dashboard/FimDashboard.tsx). Bundled IR / OSquery-ATT&CK / FIM packs; ingestsfile_eventsand synthesises alerts on writes to/etc/passwd,/etc/shadow, sshd configs, sudoers, Windows registry hives. FIM-specific detection IDsdet-endpoint-297..300. - PR6 — AiSOC osquery extensions (
services/osquery-extensions/tables/*.go). 5 custom Go-based virtual tables:aisoc_browser_extensions,aisoc_kernel_modules,aisoc_attck_persistence,aisoc_pending_actions,aisoc_alert_cache— ship richer endpoint visibility plus a bidirectional response channel.
Patch releases
v7.0.1 — Web app hardening (CodeQL + Turbopack)
- 42 CodeQL code-scanning alerts cleared (
py/unused-global-variable,py/cyclic-import,py/empty-except,py/log-injection,py/clear-text-logging-sensitive-data,py/incomplete-url-substring-sanitization,py/stack-trace-exposure,py/call/wrong-arguments,py/unused-import,js/unused-local-variable). -
apps/web/next.config.js— Removed deprecatedeslint.ignoreDuringBuildskey Next.js 16 no longer accepts; addedturbopack.rootfor workspace package resolution. -
apps/web/src/app/layout.tsx— AddedsuppressHydrationWarningto<html>so the render-blockingthemeBootstrapScriptcan writedata-theme/data-theme-preference/style.colorSchemewithout React reporting an attribute mismatch.
v7.0.2 — Version alignment + landing-page footer + docs
-
apps/web/package.jsonbumped to7.0.2; sidebar showsv7.0.2dynamically. -
apps/web/src/components/landing/Footer.tsx— Replaced hard-codedv6.1.0with a dynamic import ofpackage.json. -
README.md— Addedosquery-tls(port 8090) andosquery-extensionsto the services / Swagger / directory-tree / dev-surface tables.
v7.0.3 — Structural hydration fix + font preload
-
apps/web/src/components/layout/AppShell.tsx— Wrapped<DemoBanner />in a new<ClientOnly>boundary so the banner (which readsNEXT_PUBLIC_DEMO_MODE) is never server-rendered. Eliminates React hydration error #418 caused by stale env-var inlining producing a structural tree mismatch (server saw<button>fromSidebar, client expected<div>fromDemoBanner). -
apps/web/src/app/layout.tsx— Addedpreload: falseto theJetBrains_Mononext/font/googleconfig; eliminates "preloaded but not used within a few seconds" Chrome warnings without any visible FOUT.
v7.1.0 — Shipped ✅ (2026-05-10) — Cloud Security Coverage Wave
Six new connectors, three documentation backfills, and a new ingest template closing the biggest cloud-security gap in the connector catalogue. Every Tier-1 cloud workload protection platform now has a first-class AiSOC integration, AWS gets three native data sources, and Kubernetes audit logs land via a dual-mode connector that works on both managed and air-gapped clusters.
Track A — Documentation backfill for existing cloud connectors
-
apps/docs/docs/connectors/wiz.md— Service-account creation,read:issues+read:vulnerabilitiesscopes, token rotation, normalised severity map, worked WizIssue→ inbox event example. -
apps/docs/docs/connectors/aws-security-hub.md— IAM-role vs. static-key auth,securityhub:GetFindingspermission, and theBLOCK_IP/ALLOW_IPcapability documented end-to-end againstservices/actions/app/clients/aws_security_groups.py. -
apps/docs/docs/connectors/lacework.md— API-token flow,api_urlregional variants, alert → event severity collapse. -
apps/docs/sidebars.ts— All three backfilled pages + the four new Track B–D pages registered under theConnectorscategory.
Track B — New CNAPP connectors
-
PrismaCloudConnector(services/connectors/app/connectors/prisma_cloud.py) — Full CSPM/CWPP coverage. JWT auth viaPOST /login, paginatedGET /alert/v1/alertwith windowedtime.from/time.to, severity collapse,compute_urloverride for self-hosted Compute Edition. -
OrcaConnector(services/connectors/app/connectors/orca.py) —https://api.orcasecurity.io/api/alertswithapi_tokenauth, severity collapse with Orca-specifichazardous → highrule. - Manifests at
plugins/prisma-cloud/plugin.yamlandplugins/orca/plugin.yaml, docs atapps/docs/docs/connectors/{prisma-cloud,orca}.md, full test suites underservices/connectors/tests/test_{prisma_cloud,orca}.py.
Track C — Native AWS connectors
-
AWSGuardDutyConnector(services/connectors/app/connectors/aws_guardduty.py) — boto3-based, supports IAM-role + static-key auth via_resolve_session, iterates detectors withlist_findings+get_findings. Collapses GuardDuty's continuous numeric severity scale (0.1–10.0) into AiSOC's four-tierinfo|low|medium|highladder. -
AWSCloudTrailConnector(services/connectors/app/connectors/aws_cloudtrail.py) —cloudtrail.lookup_eventswith a curated 21-event allow-list covering identity abuse, persistence, data-plane abuse, network exposure, and trail tampering. Allow-list overridable via theevent_namesfield. -
AWSVPCFlowLogsConnector(services/connectors/app/connectors/aws_vpc_flow.py) —cloudwatch_logs.filter_log_events, v2 + v5 format parsing, RFC-5735-aware public-IP heuristic, default?REJECTfilter pattern, severity heuristic (public REJECT → medium,internal REJECT → low,ACCEPT → info). - Manifests + docs + tests for all three connectors. 27 unit tests for
AWSVPCFlowLogsConnectorcovering v2/v5 parsing and public-IP edge cases.
Track D — Kubernetes audit logs (dual-mode)
-
KubernetesAuditConnector(services/connectors/app/connectors/kubernetes_audit.py) ships with two delivery modes selected via themodeconfig field: -webhook(recommended) — apiserver pushes audit events to AiSOC's new dedicatedPOST /v1/ingest/k8s-audit/{tenant_id}route, authenticated with a shared secret in theX-AiSOC-K8s-Tokenheader (constant-time compared so a partial-prefix attacker can't shave bytes off via timing). Legacy/v1/inbox/{token}path with thek8s-audittemplate is kept as a fallback for control planes that cannot inject custom headers into the audit-webhook kubeconfig. -file_tail— AiSOC's connector pod tails a localaudit.logusing a byte-position cursor (atomically written viaos.replaceto a.aisoc-cursorsidecar) with rotation/truncation detection and a hard per-poll byte cap so a backlog can't blow up a single poll cycle. -
services/ingest/internal/handler/k8s_audit.go— Dedicated Go handler for the webhook route. Caps body size viaK8S_AUDIT_MAX_BODY_BYTES(default 16 MiB), rejects oversized batches with413so the apiserver shrinks--audit-webhook-batch-max-sizeand retries, and publishes eachEventList.items[]entry through the existing normalizer + Kafka publisher usingconnector_type: kubernetes_audit. The route is disabled (returns503) until an operator setsK8S_AUDIT_SHARED_SECRET, so a fresh install never accidentally accepts unauthenticated audit traffic. -
kubernetes_auditnormalizer profile (services/ingest/internal/normalizer/normalizer.go) — MapsauditIDtoexternal_id,verbtoactivity_name,user.usernametoactor.user.name,objectRef.{namespace,resource,name}to a compositetarget.resource.name, and translates the connector's string severity (critical|high|medium|low|info) into OCSF integer severities (5/4/3/2/1). -
k8s-auditinbox template (services/ingest/internal/normalizer/templates/k8s-audit.yaml) — maps apiserverEventpayloads onto AiSOC's normalised event shape for the legacy inbox-token path; severity is derived in the connector's_classify_severityheuristic so the same logic applies to both delivery modes. - Severity heuristic —
exec/attach/portforwardon Pod andcreateonClusterRoleBinding→high; writes toSecret/ConfigMap/ClusterRole/Role→medium; successful reads on sensitive resources →low; everything else →info. -
plugins/kubernetes-audit/plugin.yaml— 4-field config schema (mode,cluster_name,inbox_token,audit_log_path,cursor_path),category: cloud, capabilitiespull_audit+pull_alerts. -
apps/docs/docs/connectors/kubernetes-audit.md— SampleAuditPolicy+ sampleAuditSinkfor both managed and self-hosted clusters.
Cross-cutting
-
pnpm marketplace:sync—marketplace/index.json+apps/web/public/marketplace/index.jsonrebuilt; plugin count rose43 → 49. Total:total=7104 detections=6993 playbooks=62 plugins=49 mitre_techniques=493. -
apps/web/package.jsonbumped to7.1.0; sidebar and landing-page footer surface the new version dynamically.
v7.2.0 — Shipped ✅ (2026-05-13) — Stage-2 Connector + Surface Wave (feat/wazuh-connector)
Eight-commit feature wave broadening connector reach (Wazuh + auditd), making the response-action layer vendor-pluggable, replacing the NL→query template fallback with a deterministic translator, closing the threat-intel write loop to MISP, adding a blameless case post-mortem surface, and standing up a GCP Terraform skeleton equivalent to the existing AWS module.
Connectors
-
WazuhConnector(services/connectors/app/connectors/wazuh.py) — polls the Wazuh Indexerwazuh-alerts-*indices over HTTPX with basic-auth, paginates time-windowed queries, retries on 5xx with capped backoff, and normalises severity into the four-tier ladder. Marketplace manifest + per-connector docs + 24 unit tests. -
AuditdConnector(services/connectors/app/connectors/auditd.py) — file-tail of/var/log/audit/audit.logwith multi-record reassembly by msg id, hexproctitle/argvdecode, and(inode, byte_offset)cursor for log rotation. Ships withprofiles/auditd/aisoc.rulesopinionated auditctl ruleset whose-kkeys map 1:1 to detection rules; 4 new detection rules pivot offauditd_keyfor sudoers / SSH / kernel-module / systemd tampering. 444-test full connectors suite green (excludingtest_scheduler.pywhich needs theapschedulerdev dep). - Connector registry now declares 90 first-party connectors;
pnpm marketplace:syncrebuiltmarketplace/index.json+apps/web/public/marketplace/index.json.
CLI
-
aisoc plugin new <NAME> --type {enricher|connector|responder|detection|widget}replaces the old hard-codedplugin scaffoldwith per-type templates shipped inside theaisoc-cliwheel viaimportlib.resources.string.Templatesubstitution for${slug},${name},${author}; tests parameterised across all five plugin types asserting manifest validation and zero placeholder leakage.aisoc plugin scaffoldretained as alias.
Live Actions
- Generic
(vendor_id, capability)dispatcher (services/actions/app/live_actions/) — pluggableLiveActionExecutorABC + module-level registry + dispatcher with structured logging and error translation. Unknown pairs returnLiveActionResult(status=FAILED, error="executor_not_found")so the agent degrades gracefully. Adapters wrap CrowdStrike, Okta, AWS SG, Splunk so they show up asbuiltindescriptors. -
/api/v1/live-actions—discover,dispatch,dry-run. Honoursdry_run+ missing-credentials →SIMULATED, neverPARTIAL. - 45 new tests across models / registry / dispatcher / router / builtins (full actions suite: 99 passed).
-
apps/docs/docs/concepts/live-actions.md+ sidebar entry.
Agents
- Deterministic NL→ES|QL translator
(
services/agents/app/nl_query/) — IR + grammar validator + renderers for ES|QL / KQL / SPL. Replaces every# TODO: translatecomment inservices/api/app/api/v1/endpoints/nl_query.py. Optionalenhance_with_llm(gpt-4o-mini) path with deterministic fallback so the air-gapped story keeps working. - 50-pair NL→ES|QL eval set
(
services/agents/tests/eval_data/nl_query_eval.json) +test_nl_query_eval.pyharness — 100% syntactic validity, 100% semantic match (50/50 perfect) against gold intents.
API surfaces
- Blameless case post-mortem —
GET /api/v1/cases/{case_id}/postmortem?format=json|html. Pure builder + async DB orchestrator (services/api/app/services/case_postmortem.py) reusing thecase_summaryfetchers; HTML renderer with inline CSS, defensive escaping, no external assets. Tests assert XSS escaping, deterministic ordering, and explicit blamelessness (analyst handles must not surface in the narrative; assignee header line is explicitly allow-listed). - STIX → MISP push (Stage 3 #20) — closes the threat-intel write
loop.
POST /stix/indicatorsandPOST /stix/bundlesaccept?push_to_misp=trueand return a structuredmispblock on the same response.GET /stix/misp/healthandPOST /stix/misp/dry-runadded for operator verification + air-gap audits. Pure mappers coveripv4/ipv6,domain-name,url,email-addr,file:hashes(MD5/SHA-1/SHA-256/SHA-512) andfile:name. Push failures are non-fatal: AiSOC store remains source of truth, MISP is best-effort. Reuses the existingenforce_airgap_for_urlchokepoint. 76 new tests.
Infrastructure
-
infra/terraform/gcp/— Cloud Run forapi/web/ingest, Cloud SQL Postgres 16 + Memorystore Redis 7.2 on private IPs through a dedicated VPC + Serverless VPC Access connector, Secret Manager for every credential, Artifact Registry for images, one service account per Cloud Run service with least-privilegesecretAccessorbindings. Skeleton points at the public GHCR demo images so a freshapplyworks zero-config.apps/docs/docs/deployment/gcp.md+ sidebar slot betweenkubernetesandenv-vars.
Documentation
-
apps/docs/docs/operations/notifications.md— complete inventory of every notification surface in AiSOC (Web Push, Slack/Teams ChatOps, playbooknotify_slack,create_ticketsimulation, honeytoken first-touch webhooks, connector freshness alerts, on-call gating, suppression / quiet-hours, per-mechanism testing recipe). -
apps/docs/docs/plugins/lifecycle.md— operator's view of plugin states, trust modes (strict | warn | disabled), filesystem + OCI discovery, the full operator REST API with required permissions, configuration reference, upgrade/rollback semantics, and the structlog events worth alerting on. -
apps/docs/docs/integrations/misp-push.md— operator doc with config, endpoints, the STIX→MISP type table, failure modes, and the dry-run-as-air-gap-proof workflow. -
apps/docs/docs/operations/case-reports.md— covers both/summaryand/postmortemwith audience, output, automation, and runbook archive guidance. Cases summary breadcrumb now points operators at both endpoints. -
apps/docs/docs/connectors/{wazuh,auditd}.md+ per-connector sidebar entries. -
apps/docs/docs/plugins/cli.md— documents the newaisoc plugin new --typesurface. -
apps/docs/sidebars.ts— every new page registered in the correct category (Connectors, Plugin SDK, Operations, Integrations, Concepts, Deployment).
Quality
-
ruff check services/andruff format --check services/clean across the wholeservices/tree (CI scope). - Targeted lint fixes committed:
E741(ambiguousl),F402(loop var shadowingdataclasses.field),E402(post-sys.pathimport),E501long-string refactors incase_postmortem_html.py,test_misp_push.py, andauditd.py.
v8.0 — Shipped (2026-09-22)
v8.0 shipped as the close-the-loop release rather than the feature list
below. It had been reserved for the package-publish milestone, but nothing can
publish without registry credentials, so packaging moved out (see v8.1) and the
release instead wired capabilities the codebase already contained and never
called. What v8.0 actually delivered is recorded under [8.0.0] in
CHANGELOG.md and summarised in RELEASES.md.
Disposition of every item that had been listed against v8.0:
Automated IOC sharing to community MISP instances via STIX/TAXII push→ shipped in v7.2.0NL→query: "show me failed logins from new ASNs last 24h" → ES|QL / KQL→ shipped in v7.2.0 (deterministic translator + 50-pair eval set)SOC-in-a-box one-click cloud deploy (Terraform module for AWS / GCP)→ GCP module shipped in v7.2.0 (AWS already shipped)Automated retro/blameless post-mortem drafting from case timeline→ shipped in v7.2.0 (ideas backlog item promoted)- Mobile responder console (React Native) — shipped in v9.0; see the note
under v7.0 above. The "no React Native code exists in the tree" this line
used to carry is false:
apps/mobiledeclaresreact-nativeand Expo, and the capability itself had shipped earlier still as a PWA underapps/web/src/app/(responder)/. What genuinely remains is not engineering: no device build, simulator run or store submission has been performed, and the APNs and FCM credentials needed to deliver a push to a real handset are an account action, the same class of blocker as the npm and PyPI publish.apps/mobile/README.mdstates both halves. - Plugin publishing marketplace v3 (commercial plugins, revenue sharing) — open and unscheduled, and never in fact scheduled against a particular release. This line read "deferred past v8.0" through five majors, which dated the label rather than the decision; it is a scope decision, not a slipped commitment. Revenue sharing is a commercial decision rather than an engineering one, and the free packaging path is itself still blocked on registry credentials.
- MSSP RBAC enforcement on
/api/v1/actors/*(threat attribution) — shipped in v7.5.0 as part of the threat-actor attribution RBAC + port fix. - AI-generated threat intelligence briefings from public feeds — open, not scheduled.
- Embedded red-team scoring (ATT&CK coverage %) as a live dashboard widget — open, not scheduled. The underlying coverage heatmap shipped in v5.1; the dashboard widget did not.
- SLA breach predictor (ML model on historical MTTR data) — open, not scheduled.
- Incident cost estimator (breach impact calculator) — open, not scheduled.
v8.1 — Shipped (2026-09-23)
Wave-2 features plus release integrity. Every backlog item was audited against the tree before any code was written, and the backlog was wrong in both directions: two items were already built, and four had the capability present with the path that feeds it broken. Two remain partial and say so — the two attack-chain implementations still never exchange data, and ChatOps still has no proactive card push or durable approval store.
Full inventory under [8.1.0] in CHANGELOG.md. Tracked in
docs/roadmap/v8-progress.md and, for the
community-facing view, issue
#362.
Packaging is not in v8.1 either, and the reason is worth stating plainly:
release.yml already builds, packs and would upload all eight packages — the
npm and PyPI jobs are written and run on every tag. The repository's only
secret is FLY_API_TOKEN. There is no NPM_TOKEN and no PyPI trusted
publisher, so the upload steps skip with a warning by design rather than
failing the release. This is an account action, not an engineering task, and
promising it in a release that cannot perform it is the kind of claim this
project's claim-to-gate matrix exists to prevent. Packaging is therefore named
against v8.2, and it becomes a re-tag the moment the credentials exist.
v8.1.1 — Shipped (2026-09-23)
An adoption audit, and no new capability. The recurring feedback on this
project — hard to install, fabricated data, architecture nobody could follow,
documentation describing things a reader could not reproduce — turned out to
share one root cause: the documented quick start did not run the product.
./install.sh started a compose file with no ingest service, no fusion
service and AISOC_DISABLE_KAFKA: true, and the populated console a reader
saw was a seed script writing rows into Postgres.
The pipeline itself works. It had simply never been demonstrated, and now
make smoke demonstrates it on every CI run: one real event through ingest,
Kafka, fusion and detection, read back from the API, eight stages each
reporting independently.
Also in this release: /readyz on services/ingest that dials Kafka rather
than answering unconditionally, make doctor, CORE as the default ten-service
profile, five fabricated-data surfaces gated behind demo mode, the three
packages that could not be built at the v8.1.0 tag, and three new documents —
docs/audit/REPOSITORY_REALITY.md, a data-flow rewrite of
docs/architecture/README.md, and docs/testing/CLEAN_INSTALL.md.
Full inventory under [8.1.1] in CHANGELOG.md.
v12.0 — Shipped (2026-09-28)
The deployment a stranger actually starts, treated as the threat model. Three
reported vulnerabilities sat in the profile make up brings up, and none of
them needed a credential to reach: the actions service skipped authentication
entirely whenever its service token was empty and AISOC_DEV_MODE was set —
which compose defaults on while nothing generated the token — so
isolate_host, disable_user, block_ip and run_script were available to
anything that could reach the port; the realtime edge verified connection
tickets against a constant committed to this repository and read an unset
/internal/* token as authorized rather than as unconfigured; and viewer
could write to cases on all nine write routes, where a permission deliberately
withheld from the role was enforced nowhere. All three now refuse to serve
until the credential they need is set, which is what makes this a major, and
scripts/ensure_env.py backfills the generated secrets into an existing
.env so the documented path repairs itself.
Capability work in the same release: hunting stopped being a library of hunts somebody had already written, with an agent that turns a hypothesis into a plan whose fields and operators are closed enums and a compiler that binds every model-supplied value as a parameter, so injection is unrepresentable rather than filtered; the hunt corpus went from 5 to 68, and the grading over it stopped being satisfiable by accident once a negative scenario had to differ from its positive in exactly one indicator field — which found eight violations on its first run against a corpus the old grading was passing. And autonomy is now earned from a measured track record rather than typed into a setting, with promotion refusing an unmeasured rate and demotion ignoring one.
Two gates that reported success over nothing were rebuilt on what the app
actually serves: the route-shadowing check read app.routes and filtered for
APIRoute, which include_router no longer populates, so it compared zero
pairs across 456 published operations — and the defect it should have caught
had left DELETE /api/v1/autonomy-policy/grants unreachable since it shipped,
answering 204 while deleting from the wrong table.
Full inventory — 44 entries — under [12.0.0] in CHANGELOG.md.
v11.1 — Shipped (2026-09-25)
One user's bug report, and the audit it turned into. A self-hoster bringing
AiSOC up with Compose on a single host reported that it "always deploys the
demo environment, even if I add the right variables". They were right, for two
independent reasons. next build freezes both halves of the console's
routing — NEXT_PUBLIC_* values are inlined into the bundle and the
destinations returned by rewrites() are compiled into
routes-manifest.json — so a pulled image could only ever talk to the hosts
it was built against, while docker-compose.yml set the variables on the web
service where nothing read them. And every host port published to a literal
127.0.0.1, so the stack came up healthy and unreachable from anywhere but
the machine it ran on. The console's addresses are now resolved when the
container starts, and an address that is set and cannot be applied stops the
container rather than silently falling back.
A minor: nothing here requires an operator to act. Every new variable defaults to the behaviour that was previously compiled in, and the only removals are two variables the console never read.
The second root cause was that the fix would have shipped to nobody. The
console image a quickstart pulls was built from a commit two releases behind
main and aisoc-web:v11.0.0 was never pushed at all, while the rest of that
release was — every workflow green throughout, because a green workflow says a
job ran, not that the registry holds anything. The cause was arm64
cross-building under QEMU, measured at roughly 7x on a good run and
non-converging on a bad one: one arm64 pnpm install ran 6,358 seconds
without finishing while the amd64 leg of the same build took 100s. Both
publish workflows now build each architecture natively and merge the results
into a manifest list, and scripts/check_published_images.py asks the
registry whether every image the compose file, the chart and the docs name is
actually there — 14 findings against main before it existed. The Helm chart
could not have installed at all, and aisoc-honeytokens, aisoc-purple-team
and aisoc-osquery-tls are published here for the first time.
Also in this release: a recorded 2 min 57 s deployment walkthrough against the published images, with real data and real token counts. Both triage runs in it fell back to the deterministic path because the bundled 3B model's output failed schema validation — the closing card says so. No hosted provider has been exercised; there is still no funded key.
Full inventory under [11.1.0] in CHANGELOG.md.
v11.2 — Shipped (2026-09-26)
The answer to a question this repository had been asked for months: why 833
executable rules against a library of roughly 7,000? Not a missing feature —
Windows events nest their payload under System/EventData, one level below
anything the engine flattened, so CommandLine (2,173 rules) and Image
(2,300 rules) read None and no Windows rule could fire however correctly it
was written. The fix is in windows_event.normalize(), because those are
names from the Windows event schema and the engine is shared by every
connector; the engine and matcher are byte-identical.
The engine now runs 2,603 rules, and every added rule was watched to fire.
scripts/sigma_compiler.py translates imported Sigma into match_when or
refuses: of 3,132 rules considered, 1,770 ship and 1,362 are refused with a
recorded reason. The largest refusals are 556 whose log source no connector
emits and 464 whose negation would flip on a missing field, since Sigma treats
not filter as true when the field is absent and only two matcher operators
do. A rule enters the compiled ruleset only after a vendor-shaped event has
been replayed through the real connector and the real engine and produced a
hit, with an empty event of the same shape producing nothing — and
compile_sigma_ruleset.py --prove-gate reverts the connector and requires all
1,687 Windows rules to stop firing, so the proof is known to be able to fail.
It is a claim about reachability, not about detection.
A minor: nothing here requires an operator to act. No configuration
change, no migration, no API change; the connector fix only adds fields and a
normalized key still wins on collision. The detection surface is 3.1x wider
though, so a deployment with Windows telemetry should expect more alerts from
the same stream; upstream_status travels onto the alert so the 125
experimental rules can be filtered without disabling the rest.
Also here: upstream lifecycle status no longer gates execution (fireability does), all 122 previously phantom-enabled rules are resolved, windowed aggregation rules went from 10 to 18, and auto-triage asks the provider for JSON rather than correcting prose afterwards — measured 44/50 to 50/50 on the bundled local model. Four counters that still classified rules by directory were moved onto the compiled ruleset, which is what closed the long-standing disagreement between the validator, the coverage page, the marketplace index and the truth table.
Full inventory under [11.2.0] in CHANGELOG.md.
v11.0 — Shipped (2026-09-25)
What a first run actually produced. Most of this was found by bringing the
stack up from the documented path and photographing the result, and nearly
every item passed the existing test suite while failing on a real deployment.
Following the README broke the credential vault, because cp .env.example .env
wrote a placeholder the vault treats as fatal while treating empty as fine.
make up then reported the stack broken on every machine, scoring the one-shot
model pull's normal exit as a dead container. And the profile a new user runs
had no real data and no model behind it, which is also why the console's threat
page had recently been caught rendering invented indicators: the missing feed
and the fabrication were one hole.
A major because upgrading requires action on two items. POST /v1/ingest
and POST /v1/ingest/batch now require a credential and take the tenant from
it, where they previously read X-Tenant-ID and believed it — so anyone who
could reach the port could write alerts into any tenant. And CORE now needs
8 GB of memory and 20 GB of free disk, up from ~6.5 GB, because the
threat-intelligence feed, its vector store and a local model moved into it.
Re-measured rather than restated: CORE is 16 long-running services plus a
one-shot model pull and full is 22; resident memory for the whole stack went
1.72 GiB → 4.84 GiB. A fresh make up now holds 1,723 real CISA KEV entries
within a minute of boot with no credentials, and runs triage against a bundled
3B model — which returned usable triage output 44 times in 50 before its reply
was constrained to a JSON object, and 50 of 50 after, the remainder falling
back to the deterministic path and logging that they had. No hosted provider has
been exercised; there is still no funded key.
Full inventory under [11.0.0] in CHANGELOG.md.
v10.0 — Shipped (2026-09-25)
An audit of one shape: a control that exists, is tested, and sits on a path nothing reaches. Row-level security covered 92 tables and filtered nothing, because every service connected to Postgres as a superuser. Fifty-eight routes across four services carried no authentication at all, and thirty more let the caller name the tenant they were reading. UEBA reported healthy and could not write a baseline or an anomaly. AI triage was wired to a gateway that both of its resolvers had been deliberately written to ignore.
A major because upgrading requires action. Services now connect as a
DML-only aisoc_app role rather than as the schema owner, which is what makes
those 92 policies filter; an existing data volume or a managed database must be
migrated deliberately. Also breaking: the /mssp/* payloads dropped fields
that described fabricated data and now refuse a non-member, two executor-less
ActionType members were removed, CostTracker.total_cost_usd is gone, and
the published aisoc-web:latest image no longer carries demo mode — the demo
build moved to its own tag.
Two published figures were re-measured rather than restated: CORE is 11
services (the LLM gateway moved into it) and full is 21, not 30. The quick
start now ends by creating an administrator and printing a generated password
once, because the credential pair the documentation published was wrong in
three independent ways.
Full inventory under [10.0.0] in CHANGELOG.md.
v9.0 — Shipped (2026-09-23)
Ten waves of one audit question: what in this tree exists, is tested, and has no caller? v8.0 named that shape and found it a dozen times; v9.0 went looking for it deliberately and found it in the approval loop, the marketplace, the mobile console, the detection engine and the benchmark scoreboard.
The one worth stating first inverts what the feature appeared to do:
approving an action executed nothing. decide() flipped a row and never
reached the execution service, so every tap of Approve recorded a decision and
ran nothing — while telling the operator the opposite. Nothing ever created an
approval either, so the queue had no producer and was structurally empty on
every deployment.
Also: UEBA consumed a topic nothing writes; a plugin could never be rejected for a bad signature; there was no registry allow-list and no digest pinning, both of which the notes claimed existed; the public scoreboard was frozen for ten weeks with every check passing; neither published Go SDK was installable; and the Helm chart did not render at all.
What is knowingly still open is listed under ### Known in [9.0.0] —
eight PARTIAL matrix rows, the migration-on-existing-volume path, 133
unreachable detection rules, and the OCI install route held back rather than
shipped with eight unresolved high findings.
Full inventory under [9.0.0] in CHANGELOG.md.
Ideas Backlog (unscheduled)
- "Explain this alert" button using LLM with enrichment context
- Browser-extension recorder for analyst playbook capture
- Voice-driven incident commander (TTS / STT for hands-free triage)
- Automated retro/blameless post-mortem drafting from case timeline