What changes for the organisation
The wrong attachment to the wrong recipient stops being a notifiable breach.
Checks outbound mail for possible internal or accidental data disclosure before it leaves, then later the file share, the USB port and the print queue.
What changes for the organisation
The wrong attachment to the wrong recipient stops being a notifiable breach.
How it works
Body, recipients, headers, conversation context and supported attachments are analysed before delivery. Deterministic detectors and policy decide the clear cases; the private LLM judges the ambiguous ones. Outcomes are allow, warn, quarantine for authorised review, hard-block, or require approval.
The planned SMTP relay and transport-agent path is designed for customer-operated mail infrastructure. Each named platform remains subject to connector validation, capacity testing and the design-partner promotion gates stated on this page.
Where a customer operates another mail server or a non-standard routing topology, SentraMind can provide connector-customisation services to map its hold, analysis, decision, release and evidence behaviours onto that environment.
For hosted services such as Gmail, Outlook.com or Yahoo Mail, a browser extension or an Outlook client add-in can apply a similar pre-send data-disclosure check. Coverage depends on the host and client extension points available to the customer.
Postmaster looks for possible data disclosure: sensitive content sent outside policy, to an inappropriate recipient, or exposed internally to people without entitlement. Antivirus, malicious attachment and malware scanning remain with existing specialist security products.
Every step has a bounded responsibility and authority. Rules handle clear cases first, the private model judges only the remaining ambiguity, and an authorised person controls release.
Body, recipients, headers, thread context and supported attachments
The message has not crossed the governed boundary.
Active content stays disabled; the LLM never opens an attachment.
Antivirus, malicious-attachment and malware scanning remain with specialist products.
Sensitive data, Exact Data Match, document fingerprints, credentials, wrong or unentitled recipients, and lookalike domains.
The model does not receive the mailbox and cannot authorise release.
Explicit rules always win; every decision leaves a reason and evidence.
Sensitive content is not echoed to the sender; reveals and releases are attributed to authorised people.
The mechanism, not the marketing. If a step below would not survive an engineer's scrutiny, it should not be on this page.
Operating path
Not an attacker. A relationship manager in a hurry, on a Friday afternoon.
Q3_review.xlsx is attached and sent to a client contact. The message is held pre-delivery; the sender's client shows it as sending.
The workbook is extracted in the sandbox. A third worksheet, hidden and not visible in the sender's view, contains 1,340 rows belonging to a different client.
62 Exact Data Match hits against the enrolled customer table. The recipient is known-external and legitimate — but not entitled to that data.
Outcome: quarantine for authorised review. The sender sees a warning naming the category only. No content is echoed, because a sender who is about to make a mistake is not a reason to show them someone else's data.
The reviewer opens the case, sees the masked evidence under their own role, confirms it, and returns the message to the sender for correction.
One quarantine, one authorised review with the reveal logged, one release-to-sender. Nothing crossed the boundary, so no notification clock started.
What it came to
Had it delivered, this was a personal-data disclosure to an unauthorised third party. Malaysia's PDPA 72-hour clock and the client's own contractual notification terms would both have been running from the moment it was discovered — which, realistically, would have been weeks later.
The authority ladder. Green rungs are authoritative — a rule, a scanner, a sandbox, or a person. Blue is where the model contributes. Amber is where a human decides.
Judgement on the cases a rule cannot describe — the business-email-compromise pattern, the attachment that is wrong for this recipient, the thread that changed tone.
It cannot confirm a compromise or authorise a release. It flags, explains, and hands the decision to a person.
We will not show you a scripted demonstration of something that does not exist. What we publish instead is the workflow specification, held to the same standard as shipped work: actor, trigger, happy path, exceptions, terminal state, and the audit evidence it must leave. When it is built, this section becomes a demonstration.
Roadmap discovery. Promotion requires two design partners, a named mail-platform connector passing its support contract, mail-processing latency and quarantine/release accuracy targets, a capacity profile, a threat model and architecture approval.