Security & trust
Nothing up my sleeve.
Unpublished inventions are among the most sensitive documents a firm holds. This page says exactly where client data goes, who can see it, and how it is protected. Updated October 24, 2026.
For inventors
Your invention stays confidential
What you describe is available only to you and the firm you work with, and is processed only by the services named below. It is never used to train anyone's AI model, never published, and never sent to search engines or data brokers, so using ClaimMark-P does not put your invention in front of the public before your application is filed.
For firms
Built around your duties
Designed around ABA Formal Opinion 512 and the USPTO's guidance on AI tools: competence, confidentiality, supervision, and candor stay with the practitioner. Matter data stays in the United States, every subprocessor is named below, and our Security Overview answers the questions a vendor review asks.
Controls in place
Where your data lives
ClaimMark-P runs on Amazon Web Services in the United States (us-east-1), in a dedicated production account inside ClaimMark's own AWS organization. No development or test workload shares that account, and there is no demonstration data in production. One part runs outside it: the non-US patent index, described below, is on a computer ClaimMark operates itself, which is also used for ClaimMark's development and testing. It receives search text only, never matter documents, drafts, or account data.
Encryption
Every connection uses TLS 1.2 or 1.3; plain HTTP is redirected and browsers are told to use HTTPS only (HSTS). The database, document storage, application secrets, and logs are encrypted at rest with keys held in AWS Key Management Service.
Sign-in and access
ClaimMark-P accounts exist only by invitation from an administrator. Anyone can sign up for ClaimMark-A Prior Art Search, and those accounts reach search and their own organization only, never matter data. Two-step sign-in (an authenticator app) is required for administrators and every ClaimMark-A Prior Art Search account, can be required by a firm for its attorneys, and is available to every user. Repeated failed sign-ins lock the account.
Sign-in links
Invitation and password-reset links are e-mailed to the user. The e-mails that carry them are not stored, not shown to administrators, and not logged. An administrator can reveal a one-time link only after an explicit confirmation. Each reveal issues a new link, earlier ones stop working, and the account's security log records who did it, when, and why, never the link itself. Password-reset links expire after 24 hours. When an administrator changes a user's sign-in e-mail, the old address is told (with the new one masked), outstanding links stop working, and the change goes into the security log.
Firm isolation
Every matter belongs to one firm, and every request is checked on the server against the signed-in user's firm and role. A user who holds several roles acts in one role at a time.
Network boundary
The application servers accept traffic only from the load balancer, which sits behind a web application firewall. Outbound traffic from the application is limited to a short allowlist of named services; anything else is blocked at the network level. Links an inventor gives ClaimMark-P are read by a separate function outside this boundary, described below.
Monitoring and audit trail
Administrative activity in the AWS environment is recorded to a tamper-resistant audit trail kept for a year. Threat detection, configuration monitoring, and AWS security best-practice checks run continuously. Application logs are kept for 365 days, encrypted.
Backups and recovery
The database runs in two availability zones with point-in-time recovery and 35 days of backups, and is protected against deletion. A full restore was tested on October 3, 2026.
Attorney control
ClaimMark-P assists qualified practitioners. It does not give legal advice, sign, or file. Every output is kept, edited, or dismissed by counsel, and the record of how a disclosure and draft were made is kept with the matter.
Index Explorer
A system administrator can grant Index Explorer to individual users. It is not available to ClaimMark-A Prior Art Search accounts. It lets you ask an AI model (Anthropic Claude on Amazon Bedrock, listed below) questions about ClaimMark's own patent index. It can read only the public patent index, never matters or client data. Your questions, the answers, and the steps behind them are stored for you, and you can delete them; administrators see only answers you flag.
Who processes client data
The first three are the only outside services that receive matter content in production, and none of them may use it to train models; Stripe handles payments for ClaimMark-A Prior Art Search and receives no matter or search content. The application can otherwise reach only public patent sources, the USPTO and Google Patents, to fetch published documents; no confidential text is sent to them. Links an inventor gives are read by a separate function, not by the application, as the next section explains.
| Service | Purpose | Data | Location |
|---|---|---|---|
| Amazon Web Services | Hosting, storage, logging, and language-model inference through Amazon Bedrock | All matter data | United States |
| OpenAI | Language-model inference for drafting and analysis, until those models are available to us on Amazon Bedrock | Disclosure and draft content | United States |
| TypeSafe AI | Structured yes-or-no decisions on prior-art and office-action text | Excerpts of technical text | United States |
| Stripe | Payments for ClaimMark-A Prior Art Search plans, on Stripe’s own pages | Organization name, owner e-mail, billing address and card; plan and overage call counts. No matter or search content. | United States |
The US patent search index behind ClaimMark-A Prior Art Search and the prior-art step runs on ClaimMark's own servers in the same production account, not at a third party. It does not log search queries; they are processed in memory to produce results. A diagnostic error record on the index server's encrypted disk may contain part of a query that failed, and is erased when that server is replaced.
The non-US patent index (European Patent Office publications and other offices' patent families, all public data) runs on a computer ClaimMark operates itself, outside AWS, not at a third party. ClaimMark-A searches and attorneys' prior-art searches send it their search text: for a prior-art search, the invention summary and independent claims. It never receives uploaded documents, drafts, or account data, and inventor-stage searches do not use it. The computer opens the connection itself, to a relay inside our production account that only the application can reach; it accepts no connection from the internet, and no third party carries the traffic. The search text travels encrypted end to end, from the application to that computer, with its own certificate and a read-only login, so the relay cannot read it. It does not log search queries; a diagnostic error record on that computer's disk may contain part of a query that failed. When the computer cannot answer in time, the search returns its US results and says the non-US results were unavailable.
Amazon Bedrock does not store prompts or outputs. OpenAI's standard API terms allow it to keep requests for up to 30 days for abuse monitoring; we have asked OpenAI and TypeSafe for zero-retention terms and will update this page when they are in place. Uploaded documents are read on our own servers, and no outside service sees them during text extraction.
Links an inventor gives ClaimMark-P are read outside our network
When an inventor gives ClaimMark-P a link, the application never contacts that website itself. Its outbound allowlist is unchanged, and every other outbound name stays blocked for the application. The link is read by a separate, single-purpose AWS function outside ClaimMark's private network. That function has no route to the database, document storage, the patent index, or secrets. It has no credentials of its own, only permission to write its own encrypted log.
- It reads only public http and https addresses on the standard ports. It looks up the address itself, refuses private, internal, loopback, link-local, cloud-metadata, and other reserved addresses (including disguised forms of them), and connects only to the address it checked.
- It follows at most three redirects, checks each one again, and refuses a secure page that redirects to an insecure one.
- It sends no cookies, passwords, or sign-in headers.
- It reads at most 10 MB and gives up after 30 seconds.
- It accepts only web pages, plain text, Markdown, PDF, and GitHub's public API, and it never runs scripts.
- It logs only the site's host name, the result, the size, the content type, and the time taken; never the page path, the query, or the content.
- Each inventor can add 20 links an hour, and each firm 100.
What comes back is labelled untrusted and treated as information about the invention, never as instructions. The websites are destinations the inventor chose, not subprocessors, so the table above does not change.
SOC 2
Our security controls are designed to the AICPA SOC 2 Trust Services Criteria (Security and Confidentiality). We plan to begin our SOC 2 Type II observation period in the first quarter of 2027 and will share the report under NDA once it is issued. SOC 2 is an independent auditor's attestation, not a certification, and we describe it that way.
Security Overview on request
The detailed overview covers architecture and data flow, each control and how it is operated, the subprocessor list with training and retention terms, and our SOC 2 plan. Firms receive it for vendor review, under NDA where required.