The AI that judges your email never sees it.
Isolated tools read each message and hand our reasoning engine findings, not content. Clean email is discarded the moment it is cleared. Your email never trains a model.
- SOC 2
- AWS is our only sub-processor
- US regions only
Isolated tool runtime
Raw content stays here. Never logged, never backed up, never passed on.
Reasoning engine sees
- Our difference
0 bytes Raw email in the AI’s contextThe most capable model sees the least data: findings only, never content.
-
0 Models trained on your emailFixed model weights. No fine-tuning, ever.
-
0 s Clean email retainedNo logs, backups or snapshots of content.
-
1 Sub-processorAWS only, US regions. Model providers get no runtime access.
-
90 days Certified purge on exitWritten certification. No remnants.
Each of these is enforced by how the system is built, not by policy. The walk-through below shows how.
How an email moves through the system, and exits.
Five steps, one branching decision: what happens at each stage, and what does not.
-
Step 01 Arrival
Happens
Email arrives via TLS 1.3 and is held in encrypted temporary storage within AWS US infrastructure, isolated per tenant, for the duration of analysis.
Does not happen
No copies are made for later analysis. Email bodies are not logged. Content is excluded from backups, snapshots, and disaster recovery processes by design.
-
Step 02 Evidence extraction
Happens
A set of specialized analytical tools processes the raw content in isolated per-tool runtimes. Each tool returns structured findings, for example “document structure consistent with PDF invoice, PII pattern indicators present, anomaly score elevated”, to an internal handoff layer.
Does not happen
Tools do not share raw content with each other. Raw content is not persisted beyond the tool’s isolated runtime. The number and internal functions of the tools are intentionally not enumerated on this page.
-
Step 03 Reasoning
Happens
The primary AI reasoning engine receives only structured findings from the evidence layer. It evaluates cross-signal correlation, context, and final verdict using in-context learning with fixed model weights.
Does not happen
The reasoning engine never receives raw email bodies, attachment contents, PII strings, or customer identifiers. No fine-tuning, no gradient updates, no production learning from customer data. Model providers have no access to runtime prompts or completions.
-
Step 04 Verdict
Happens
The system renders a per-message verdict before the email reaches the user’s mailbox. Decisions are auditable within the customer tenant and carry the evidence that supported them.
Does not happen
Decisions are not reused across customers without de-identification. Content does not enter any shared pool at this stage. Nothing beyond the verdict and its structured evidence leaves the tenant boundary.
-
Step 05 Fork and exit
If clean
Non-malicious email content is discarded immediately. No logs, no backups, no snapshots, no retention. Because there are no persistent write paths for non-malicious content, accidental retention is structurally impossible, not just policy-prohibited.
If malicious
Security indicators (URLs, domains, hashes, infrastructure patterns) are extracted and pushed through the de-identification pipeline. Customer-attributable metadata is stripped. Timestamps are rounded to 24-hour windows. Indicators unique to fewer than ten organizations are suppressed or generalized. Only de-identified artifacts enter the cross-customer Threat Knowledge Base, validated annually by re-identification adversary testing.
Five principles, structurally enforced.
Each principle is a specific architectural decision enforced by the system’s design, not dependent on operational discipline or policy compliance alone.
-
Principle 01
Zero-retention processing
Non-malicious email content is held only during analysis and discarded immediately upon clean verdict.
Why a CISO cares
The shortest possible retention window is the smallest possible exposure surface. Content that does not persist cannot be breached, subpoenaed, or leaked.
Controls that enforce it
- No write paths for non-malicious content into persistent storage
- Email content excluded from backups, snapshots, disaster recovery
- No logging of email bodies at any analysis stage
- Memory not snapshotted for diagnostic purposes
- Automated audits validate no leakage into logs or artifacts
Evidence in the Trust Center
-
Principle 02
Evidence-based reasoning
The reasoning engine operates on structured findings, not raw content. The most powerful model sees the least raw data.
Why a CISO cares
AI reasoning models are the most capable and the most sensitive. Keeping raw content out of their context window eliminates the largest class of AI privacy incidents by architecture, not policy.
Controls that enforce it
- Specialized analytical tools process raw content in isolated environments
- Tools return structured findings (scores, flags, indicators), not content
- The reasoning engine receives only structured findings
- No shared data store or shared compute between evidence tools and reasoning
Evidence in the Trust Center
-
Principle 03
Model isolation
No customer email content is used to train, fine-tune, or adapt foundation models. Reasoning uses in-context learning with fixed model weights.
Why a CISO cares
Training data can be reconstructed from a model’s outputs. A model never trained on your data cannot leak your data. Zero training is zero reconstruction risk.
Controls that enforce it
- In-context learning only, no gradient updates in production
- Model weights fixed and versioned, no production fine-tuning
- Supporting ML components train only on structured artifacts, never on email bodies or PII
- Training environment physically isolated from production inference pipeline
Evidence in the Trust Center
-
Principle 04
Intelligence without attribution
Cross-customer threat intelligence is derived through a multi-control de-identification pipeline designed to resist re-identification.
Why a CISO cares
Collective threat intelligence is only valuable if it cannot be traced back to any single customer. A leaky IOC pipeline turns your vendor into your breach.
Controls that enforce it
- Only de-identified security indicators enter cross-customer intelligence
- Timestamps rounded to 24-hour windows
- Behavioral indicators aggregated across a minimum of three independent sources before release
- Indicators unique to fewer than ten organizations suppressed or generalized
- Customer domains, internal identifiers, employee names stripped automatically
- Annual re-identification adversary testing validates the pipeline
Evidence in the Trust Center
-
Principle 05
Infrastructure-enforced privacy
Data protection is enforced by cloud infrastructure architecture itself, not solely by operational policy. Model providers have zero runtime access to customer data.
Why a CISO cares
Policy controls can be bypassed by people, processes, or mistakes. Infrastructure controls cannot. The strongest privacy guarantees are the ones that make the wrong behavior impossible, not just forbidden.
Controls that enforce it
- AI inference runs through AWS Bedrock in AWS-owned Model Deployment Accounts
- Model providers deliver weights only, no access to customer prompts or completions at runtime
- Sole sub-processor is AWS. All data stays in US regions
- Runtime data cannot traverse to model provider infrastructure
- 90-day post-termination purge with written certification
Evidence in the Trust Center
What the reasoning engine sees, and what it does not.
The AI reasoning layer receives structured findings from the evidence layer. Raw content, raw identifiers, and raw user-attributable data never enter its context window. Evidence-based reasoning is an architectural choice, not a runtime guardrail.
What the reasoning engine sees
Structured findings. Scores. Flags.
- URL verdict: safe, suspicious, or maliciousDerived by a dedicated URL analysis tool that walked the redirect chain in isolation
- Domain reputation scoreFrom infrastructure intelligence, not from content
- Document type and structural hintse.g. “Excel file, has macros, 50 rows of structured data”
- PII pattern indicators (present or absent)Flag only, never the actual SSN, card number, or name
- Behavioral anomaly scoreRelative to tenant baseline, numeric only
- Sender infrastructure fingerprint classPattern category, not the raw headers
- Attachment hash class (known good, unknown, known bad)
- Link redirect depth and terminal-page properties
- Language classification and tone indicators
What it never sees
Raw content. Raw identifiers. Raw bytes.
- Raw email body textNot even a summary of it
- Attachment bytes or document contentsStructural hints only, never the data inside
- Actual PII stringsSSNs, card numbers, personal names are flags, not values
- Customer internal domain names
- Employee names or email addresses
- Raw HTTP responses from followed URLsThe URL tool returns a verdict, not the page HTML
- Image contents beyond OCR-extracted structural features
- Internal customer identifiers
- Message metadata tying content to specific senders or recipients
The most powerful model sees the least raw data.
Compliance posture, in one place.
This page describes what makes StrongestLayer’s AI privacy architecture different. Compliance attestations are table stakes, not differentiators, and they live where table stakes belong: on our always-current trust center. The controls described on this page are validated within our active SOC 2 audit scope.
Trust Center, powered by Vanta
Live attestations. Always current. No email required.
SOC 2 attestation, GDPR-compliant data processing, sub-processor list, DPA, security questionnaires, BAA available on request, and audit report summaries in one continuously updated place.
Visit the trust center →Want a deeper compliance briefing?
Our sales team can walk through the full architecture with your privacy and security leads.
Schedule a briefing →Based on the Detecting Without Exposing whitepaper, StrongestLayer Security Architecture, February 2026.
FAQ
Privacy questions, answered.
Does StrongestLayer train AI models on customer email?
No. No customer email content is used to train, fine-tune, or adapt foundation models. Reasoning uses in-context learning with fixed, versioned model weights, and supporting ML components train only on structured artifacts, never on email bodies or PII.
How long does StrongestLayer keep my email?
Non-malicious email is held in encrypted temporary storage only while it is analyzed, then discarded the moment it is cleared: no logs, backups or snapshots of its content. When a customer leaves, all customer data is purged within 90 days with written certification.
What does the AI reasoning engine see?
Only structured findings from isolated analysis tools: scores, flags and indicators such as a URL verdict, a domain reputation score or a PII “present or absent” flag. It never sees raw email bodies, attachment contents, actual PII strings, employee names or customer identifiers.
Where is my data processed, and who are the sub-processors?
AI inference runs through AWS Bedrock in AWS-owned Model Deployment Accounts, and all data stays in US regions. AWS is the sole sub-processor. Model providers deliver model weights only and have no access to customer prompts or completions at runtime.
How is cross-customer threat intelligence kept anonymous?
Only de-identified security indicators enter it. Customer domains, identifiers and employee names are stripped, timestamps are rounded to 24-hour windows, indicators unique to fewer than ten organizations are suppressed or generalized, and annual re-identification testing validates the pipeline.
Where can I find SOC 2, the DPA and a BAA?
In the StrongestLayer Trust Center, powered by Vanta: SOC 2 attestation, GDPR-compliant data processing, the sub-processor list, the DPA, security questionnaires and audit report summaries, with a BAA available on request. No email is required to view it.
Do you have an AI gap? Find out what’s getting through.
See which generations of defense you’re running and what it takes to close the gap. Two minutes. No signup. Nothing changes in your environment.