Product
Solutions
Resources
Customers
Company
Platform Architecture

Distributed Detection Model

How a message travels through Sublime

From first contact to verdict — every analysis layer, detection rule, response action, and AI agent. Click any node to explore.

Core platform & rules
On-demand engines & enrichments
Attack score
Actions
ASA / ADÉ agents
📨
Email arrives
Inbound · Internal · Outbound

Sublime protects all three email traffic types through your existing mail flow. No rerouting required — integrates via API or inline MTA with Microsoft 365, Google Workspace, and IMAP sources.

Microsoft 365Google WorkspaceIMAPAPI ingest
🔍
Message ingestion & parsing
Headers · body · attachments · URLs

Every message is fully decomposed before analysis. Attachments are recursively unpacked, QR codes decoded, embedded images extracted for OCR, and all URLs pulled for live navigation. The result is a rich Message Data Model (MDM) that detection rules query against.

Recursive unpackingQR decodeOCR extractionMDM schema
MQL Detection Rules — evaluated on every message
📡
Sublime Core Feed
Curated open-source rules
A live, community-sourced rule feed covering BEC, brand impersonation, vendor compromise, homoglyph domains, URL shorteners, QR phishing, malware, sextortion, callback phishing, and more. Human-readable MQL, available on GitHub.
Open-sourceLive updatesGitHub repo
✏️
Custom detections
ADÉ-authored + analyst-authored
ADÉ autonomously authors new detections tailored to your environment. Analysts write org-specific rules in the built-in MQL editor. Query any field in the Message Data Model. YARA rules for attachments supported. No vendor ticket ever required.
// HTML smuggling via attachment name: HTML smuggling via attachment severity: high source: | type.inbound and any(attachments, .file_extension in~ ('htm', 'html') )
ADÉ-authoredAnalyst-authoredNo ticketsFull MDM access
↓ Rules call on-demand
On-demand engines
🧠 NLU engine
Intent + tone analysis
NLU moves beyond keyword matching to reason about the full intent of a message — detecting social engineering tactics, urgency cues, impersonation patterns, and fraud signals in conversational context. Called by rules that need semantic understanding.
📊 Behavioral signals
Sender + org-specific context
Sender reputation, communication history, infrastructure analysis, and org-specific behavioral baselines flag lookalike domains, thread hijacking, and abuse of trusted cloud services. Called by rules that need context beyond the message itself.
Enrichments
👁 Computer vision
Brand impersonation detection
Screenshots, embedded images, and linked pages are visually analyzed to catch brand impersonation — logos, login pages, and UI patterns that mimic trusted services. Called by rules that need to verify visual content.
Computed in parallel
Detection verdict
Rule matches · full rationale

Unlike black-box models, every Sublime verdict is backed by the specific rules that fired and clear evidence of why. Analysts see exactly which rules matched and what signals triggered them — enabling instant remediation without a vendor ticket.

ExplainableMQL rulesFull rationaleNo black box
🎯
Attack score
Fired rules + insights → model

Attack score is computed in parallel with the detection verdict — not after it. It takes the set of fired rules and identified messaging insights, runs them through a dedicated model, and produces a numeric confidence signal for prioritizing response across a campaign.

Fired rules inputMessaging insightsSeverity signalParallel to verdict
⚙️
Actions
Configured per-rule · fire automatically on match

Each detection rule triggers one or more actions automatically on match — no manual intervention required unless you want it.

🔒 Quarantine
Makes the message inaccessible. In M365, moved to Recoverable Items (restorable 14 days). In Google Workspace, deleted but restorable by admins. Also removes calendar events.
🗑 Trash
Moves to user's trash. Less disruptive than quarantine — appropriate for lower-confidence signals where user accessibility may still be needed.
⚠️ Warning banner
Inserts a visible warning before delivery. Useful for ambiguous messages where user judgment is appropriate — flagged, not blocked.
📤 Move to spam
Routes to spam/junk. Lighter-touch than quarantine, for nuisance or low-confidence signals.
🔔 Alert
Fires an email or Slack alert when a rule matches. Optional EML attachment for forensics. Configurable per rule and severity.
🔗 Webhook
Sends a structured payload to any downstream system — Tines, Splunk, custom SOAR playbooks.
🚫
Threat detected
Quarantine / trash / banner
Configured action executes automatically. Analysts see matched rules, evidence, and full message context — release a FP, scope an exception, or extend a block in one click. No vendor ticket.
Auto-quarantineInstant alertAnalyst control
🚩
Flagged for review
Warning banner · ASA queue
Ambiguous messages get a warning banner before delivery, and are routed to ASA's triage queue. Investigated and resolved in seconds — no analyst time required unless escalation is warranted.
Warning bannerAuto-triageSeconds not hours
Clean message
Delivered to inbox
Legitimate messages pass through without disruption. Org-specific detections mean significantly lower false positive rates than centralized models.
Low FP rateNo disruption
No rules fire → delivered
ASA
Autonomous Security Analyst
Triages, investigates, and resolves user-reported email in seconds — clearing the abuse mailbox automatically.
ASA handles the entire user-report workflow autonomously: reads the message, cross-references campaigns, applies detection logic, and either resolves or surfaces a full dossier to an analyst. Hours of daily triage collapses to seconds. Signal also feeds back to ADÉ.
Auto-triageAbuse mailboxSignal → ADÉ
ADÉ
Autonomous Detection Engineer
Analyzes missed attacks, writes new MQL detection rules, backtests them, and surfaces them for one-click approval.
ADÉ operates like a detection engineer embedded in your org. When a missed attack is identified, it analyzes patterns, reviews existing rules, authors tailored MQL, and backtests. Coverage gaps that used to take months now close in hours.
Missed attack → analysis → rule authored → backtested → analyst approves → live → feeds back into rules layer
Pattern analysisMQL authoringBacktestingEnterprise beta
Business Value

Autonomous by default · control on demand

How Sublime delivers value at every step

From first contact to closed coverage gap — every layer stops more attacks, eliminates busywork, and adapts at adversary speed. Click any node to see the business outcome.

Core platform & rules
On-demand engines
Attack score
Actions
AI agents (ASA & ADÉ)
📨
Email arrives
Inbound · Internal · Outbound

Most vendors only protect one direction. Sublime covers inbound, internal, and outbound email through your existing mail flow — preventing data loss and policy violations with no rerouting required.

InboundInternalOutboundM365 · Google · IMAP
Users can flag suspicious mail straight from their inbox — it routes directly to ASA for instant triage instead of sitting in an analyst queue.
🔍
Message ingestion & parsing
Headers · body · attachments · URLs

Attackers hide payloads in formats your team doesn't have time to inspect. Sublime decomposes every message before analysis — unpacks nested attachments, decodes QR codes, reads embedded images via OCR, and pulls every URL for live inspection. Nothing gets through unchecked.

Recursive unpackingQR phishingOCR-extracted textEvery URL extracted
MQL Detection Rules — evaluated on every message
📡
Sublime Core Feed
Curated open-source rules
Strong baseline coverage from day one. Sublime's research team and open-source community ship live detection updates covering BEC, brand impersonation, vendor compromise, homoglyph domains, URL shorteners, QR phishing, malware, callback phishing, and more.
Day-one efficacyLive updatesOpen-source
✏️
Custom detections
ADÉ-authored + analyst-authored
ADÉ autonomously authors new detections tailored to your environment. When your team wants to step in, the built-in editor lets you author, scope, or tune any detection directly. No vendor ticket, no release cycle.
// HTML smuggling via attachment name: HTML smuggling via attachment severity: high source: | type.inbound and any(attachments, .file_extension in~ ('htm', 'html') )
ADÉ-authoredAnalyst-authoredNo vendor ticketsOrg-specific
↓ Rules call on-demand
On-demand engines
🧠 NLU engine
Intent + tone analysis
Catches BEC, vendor compromise, and intent-driven attacks that legacy keyword filters miss. Reasons about urgency cues, tone manipulation, and executive impersonation in full conversational context.
📊 Behavioral signals
Sender + org-specific context
Catches lookalike domains, thread hijacking, and abuse of trusted cloud services that one-size-fits-all models can't see. Uses your org's actual communication history and infrastructure baselines.
Enrichments
👁 Computer vision
Brand impersonation detection
Catches credential theft attempts that hide in pixels, not text. Inspects logos, login pages, and screenshots — impersonated brands and fake portals blocked before users click.
Computed in parallel
Detection verdict
Rule matches · full rationale

Every verdict shows the exact detection that matched and the signals that triggered it. Analysts make confident calls, defend decisions to leadership, and remediate in minutes — not after opening a vendor ticket.

TransparentAuditableNo black boxConfidence at scale
🎯
Attack score
Fired rules + insights → model

When multiple detections fire across a campaign, attack score gives your team a clear severity signal. No guessing which alerts matter — the highest-risk threats reach the top of the queue.

Faster prioritizationCuts alert noiseCampaign-awareTriage in minutes
⚙️
Actions
Configured per-rule · fire automatically on match

Threats are quarantined, banner-flagged, or routed to spam automatically — no daily review queue, no human in the middle of every decision. Your team intervenes when they want to, not because they have to.

🔒 Quarantine
Pulled from inbox before the user sees it. Restorable by admins in M365 (14 days default) and by Sublime admins in Google Workspace. Calendar invites removed too.
🗑 Trash
Lower-friction option for lower-confidence threats. The user keeps recoverability without the inbox interruption.
⚠️ Warning banner
Delivered with a visible alert when human judgment is appropriate. Better than blocking ambiguous mail outright.
📤 Move to spam
Light-touch routing for nuisance signals. Keeps legitimate-but-low-priority mail recoverable.
🔔 Alert
Pings your inbox or Slack when a high-severity detection fires. Optional EML for forensics. Configurable per detection.
🔗 Webhook
Triggers Tines, Splunk, or custom SOAR playbooks directly from a Sublime detection.
🚫
Threat detected
Action fires · message blocked
Your team sees what fired, what evidence triggered it, and which mailboxes it touched. Release a FP, extend a block to a wider campaign without filing a ticket.
Auto-blockedInstant alertOne-click control
🔎
Flagged for review
Suspicious · unknown · error → ASA
Suspicious verdicts, unknowns, errors, and user-reported messages all route to ASA. Your team gets a triaged dossier — or nothing at all when ASA resolves it autonomously.
80% faster triageCleared queueUser-reported handled
ASA
Autonomous Security Analyst
Triages flagged messages and user-reported email in seconds — clearing the queue automatically.
ASA reads flagged and user-reported messages, cross-references known campaigns, applies detection logic, and either resolves or surfaces a complete dossier. If a threat is found, it triggers the action — quarantine, trash, or banner — without anyone in the loop. Customers see 80% faster user-report investigation.
80% faster triageAlways-onAction-capableSignal → ADÉ
ADÉ
Autonomous Detection Engineer
Activated by ASA when coverage gaps are found — writes new MQL detections, backtests, and publishes them.
When ASA surfaces a gap, ADÉ analyzes patterns, reviews your existing detections, authors new ones tailored to your environment, and backtests against historical traffic. Coverage that used to take a vendor cycle now ships in hours, on your terms.
ASA finds gap → hands to ADÉ → patterns analyzed → tailored detection authored → backtested → published → folds into your detection set
Autonomous by default, control on demand. ADÉ can publish without analyst approval (opt-in) or surface every detection for one-click review.
Hours, not monthsTailored coverageBacktestedAuto-publish (opt-in)