◈
Findings ledger
Every verdict becomes a durable, idempotent row with evidence references and scorer version pinned for replay.
defender-side confidence engine
Your attackers are attacking at machine speed.
Why is your team defending at human speed?
Security decisions with evidence attached.
AEGIS turns noisy findings into calibrated verdicts—grounded in scope, evidence, priors you own, and a ledger that never silently drops a decision.
Built for defenders who need automation without surrendering explainability. Every action stays inside owned-asset boundaries, every score can be replayed, and uncertainty remains visible.
I1
Never silently drop
I2
Criticality-aware bands
I3
Versioned scorer
I4
Thinness stays visible
The defender's problem
Signal is cheap. Trust is not.
Security teams receive more findings than they can investigate, while simple automation often hides the assumptions that make a verdict unsafe.
AEGIS is the decision layer between detection and action: it audits scope, gathers bounded evidence, combines it with calibrated priors, and records the result for later review.
Live platform sandbox
From alert to accountable verdict.
Choose a finding class and watch the AEGIS pipeline make uncertainty, scope, evidence, and disposition visible in one control center.
Execution stream● LIVE
TRIGGER
GuardDuty: unusual AssumeRole activity on production account
[09:14:02] ALERT INGEST
Finding received with asset scope: prod-accounts/*
scope: lockedledger: onv1.0
⟁
AEGIS Orchestrator
evidence loop activecollector
Finding received with asset scope: prod-accounts/*
No action executes without scope and disposition checks.
Verdict pendinginvestigating
Investigation in progress
The verdict card will appear after the evidence loop and ledger write are complete.
1 / 6 stages complete
interval: pendingcalibration: pending
The decision layer
Automation with its assumptions intact.
AEGIS makes the invisible parts of security automation inspectable: scope, evidence quality, data thinness, model influence, and what happened after the verdict.
∿
Empirical-Bayes priors
Feedback separates two questions a single label conflates: was the detection correct, and did it warrant action. Thin data never masquerades as certainty.
⊙
Calibrated scoring
Soft evidence is clamped; hard signals are explicit. The engine exposes the reasoning path without leaking thresholds.
△
Cost-derived bands
HIGH, BOUNDARY and LOW come from your cost model, not ours: you set what a missed finding and a wasted hour are worth, and the thresholds follow per asset.
↻
Auto-close that comes back
Closed findings re-enter a retest pool and the floor drifts, so activity shaped to sit just under the threshold cannot stay there. This is what makes auto-close defensible rather than merely quiet.
⋈
Attack-path chaining
Findings that are individually low and jointly high are scored as a path, not a list. A chain that reaches a critical objective lifts every step in it.
⌁
Scope enforcement
Pre-tool hooks keep collection inside owned assets and allowlisted operations before evidence is gathered.
✣
MCP-ready operations
Playbooks, scoring, feedback, and explanation are exposed as composable tools for defender workflows.
Integration surface
Bring the signal. Keep control of the response.
SIEMs and cloud audit logs send findings in. Approved actions go back out as typed, scoped intents, and the connector that enforces them can run in your infrastructure holding your credentials, so keeping control is an architectural fact rather than a promise. A decision layer should not need god-mode keys to your identity plane.
Upstream intake
ALERTS INSIEMs and detectors
Google Cloud audit logs
Cloud provider · liveA Pub/Sub forwarder deployed inside the customer project maps audit events to the intake contract. Running in production today.
Wazuh
Open-source SIEMCustom Manager Integrator sends structured alerts into the AEGIS webhook for host, FIM, vulnerability, and security monitoring.
Splunk
Enterprise SIEMSaved searches and alert actions feed correlated findings into the same normalized intake contract.
Elastic / OpenSearch
Detection analyticsAlerting webhooks and forwarding adapters connect search-based detections to the evidence loop.
Microsoft Sentinel
Cloud SIEMLogic Apps and automation rules route Microsoft security findings into AEGIS for calibrated disposition.
Wazuh path: use the open-source Manager Integrator with JSON alerts and an authenticated AEGIS webhook. Stable source IDs preserve deduplication and replay.
Downstream actions
ACTIONS OUTResponse connectors
isolate_asset
availableAWS EC2 security groups
plannedCrowdStrike Falcon · SentinelOne · Kubernetes NetworkPolicy
revoke_session
availableRedis token blacklist
plannedOkta · Microsoft Entra ID
disable_identity
availableAWS IAM
plannedMicrosoft Graph / Entra ID · Okta
block_network_indicator
availableCloudflare WAF · AWS WAFv2
plannedNetwork firewall and IP-set services
Approval is a grant, not a click per alert. A human authorises a class of action on a class of asset for a bounded window; each execution cites that grant, carries an idempotency key, and writes its provider result back to the ledger. Approving one action at a time does not survive volume: it produces consent fatigue and an audit trail that records agreement without attention.
Architecture
A narrow, auditable path from signal to action.
AEGIS keeps the model in its lane. The engine owns the ledger, priors, scoring, and bands; hooks own scope; playbooks define the evidence workflow; MCP exposes the capability without making the model the source of truth.
01
Finding
raw signal
02
Scope
owned asset
03
Evidence
bounded proof
04
Prior
feedback
05
Score
log-odds
06
Band
route
07
Ledger
replay
Start with one workflow
Request an AEGIS pilot.
Bring one noisy finding, one owned environment, and one decision you want to make safer. We will help scope an isolated pilot around it.
Pilot requests are reviewed manually. Please do not include secrets, credentials, or sensitive production data.