Security

Secure by construction.

Most of our work runs on public record. The rest stays locked to your company, behind human review.

Read the plain answers

Where your files go

Your browserSigned in to your company only

Krypton AI applicationNothing reaches you unreviewed

Private perimeter

Database and storageAES-256 on every disk and backup

Row-level security per company
Embargoed files cannot be indexed
Separate research and drafting roles
Append-only audit

Encrypted at rest

AES-256 on every disk and backup.

Encrypted in transit

TLS on every connection.

Company-scoped

Row-level security keeps each company apart.

Embargo lock

Non-public files cannot be indexed or embedded.

Human review

Nothing reaches you unreviewed.

Append-only audit

Grants, exports and edits leave a record.

Multi-factor access

Outreach staff need MFA and a grant per company.

Separate roles

Research and drafting run under separate database roles.

Public record only

Most of the work never touches your files.

Peer watchOther companies' public filings.
Ownership13F, 13D, 13G and Form 4.
Investor targetingPublic data about the investor.
DistributionMaterial already published.
Sell-side introductionsNo company data changes hands.

Private perimeter

What does, stays inside it.

Earnings releases before the wireNon-public until issued.
Decks ahead of a quarterNon-public until disclosed.
Post-meeting debriefsCan carry what a fund heard.

Separated by the database role model, not by policy.

The plain answers

Updated 29 October 2026

Krypton AI does not hold SOC 2 or ISO certification. Read the public-record services in the standing routines.

Every data hop: Browser to the hosted application to a private database and storage.
Non-public material: Uploads begin unverified or embargoed. Database constraints prevent embargoed documents from being chunked or embedded, and research and drafting roles are separate.
Krypton AI staff access: Client membership is company-scoped. Outreach requires multi-factor authentication, current attestations and an explicit company-and-audience grant. Verified operators manage access.
Audit trail: Grant and revocation, brief publication and viewing, claim copying, activity and correction, exports and provider versions create append-only audit evidence. The system does not claim that every read or administrative action is logged.
Tenant isolation: Clients share one database project and are separated logically by company identifiers, row-level security, tenant-bound sessions and database constraints. Each client does not have a dedicated database, schema or encryption key.
Encryption at rest and in transit: Our hosting provider states AES-256 encryption at rest and TLS 1.3 in transit. Our database provider states AES-256 encryption for disks and backups and TLS 1.2 for network traffic. Recall.ai, which holds meeting recordings and the calendar connection, states that bot data is encrypted at rest; it publishes no algorithm, and this has not been verified against its audit report. Krypton AI does not add application-layer or customer-managed encryption keys.
Retention and deletion: Core documents and audit records do not have a general time-based expiry, and a complete tenant purge has not been implemented or tested. Outreach evidence is append-only. Krypton AI does not claim deletion on request.
Follow-up drafts: Follow-up drafts after a meeting are written by an AI model from the meeting record and your notes, behind their own switch. Our team approves each one, and only you send it.
Briefing PDFs: A briefing on a fund or analyst is made when you ask, from your own outreach record and public filings. Its likely questions are written by an AI model and checked so every number in them is one of the briefing's figures. Each briefing made is recorded in your audit log.
Signal storylines and claims: Signal groups public social media posts about a company and its peers into storylines, named by an AI model, and checks claims in those posts against the company's public SEC filings. Only public posts and public filings are sent to a model, never client material. Storylines and checked claims appear without review first, our team can remove any claim, and the whole feature runs behind its own switch.
Signal sentiment: Signal's sentiment score is read by an AI model from public posts about your company and its peers. Only each post's words, its source, its author's public handle and its date are sent, never your material or your contacts. Your company's own posts are not scored.
Signal news and corrections: Signal also reads financial news about a company and its peers from free public sources (Stock Analysis, Yahoo Finance, SEC 8-K filings and the newsletters we follow) and keeps each story's headline, first lines, link and publisher. The claim check also reads the company's own press releases and its own public posts, and a verdict that rests on them says so. When you mark a post's score or a storyline's tone wrong, the post's words and the side you gave are sent to the AI model as examples when it scores other posts, never who made the correction or for which client.
Signal monthly reports: On the first business day of each month Signal makes a one or two page PDF of the month before for each client, from public posts and the claims checked against public SEC filings. Its summary paragraph is written by an AI model from the report's computed figures only and checked so every number in it is one of them. The PDF is kept in private storage, shown only to the company's members after the second sign-in step, and a member who turned on Signal alerts gets an email with a download link that lasts 14 days. The summary runs behind its own switch.
Signal response drafts: For the storylines about a client that matter, an AI model drafts posts for each platform, an investor FAQ entry, a press release draft, an earnings call prep pack and posts in the chief executive's voice, and matches the client's latest SEC release to what public posts say. Only public posts, public SEC filings and the company's own published releases, call transcripts and decks are sent to the model, never private uploads or our records. Every figure is checked against its source. Drafts appear without review first, nothing is ever posted or sent for the client, our team can remove any draft, and the feature runs behind its own switch.
Subprocessors: The core path uses our hosting and database providers. Email, calendar and meeting workflows can additionally use Postmark, Google or Microsoft, Recall.ai and Resend.
Microsoft environment compatibility: Microsoft Outlook calendar connection and browser access are supported. Entra single sign-on, SharePoint or OneDrive content integration, Azure or virtual-network deployment, customer-managed keys and processing inside a customer's Microsoft tenant are not implemented.