Why you don’t have to take our word for it

A leak shouldn’t require blind faith. So everything on this page is concrete and verifiable — how the site is built, what we will and won’t do, and exactly where the limits are — checkable against our open source rather than simply asserted.

And you aren’t handing material to an anonymous form. CUPIDS is a named, accountable lab at CU Boulder with a public-interest mission: training students to preserve at-risk public data and to back the journalists, lawyers, and nonprofits who hold power to account. That work collapses if the people who come to us aren’t safe — so protecting them is a design requirement here, not a footnote. Our incentives are to protect you, our methods are open to inspection, and where we can’t promise something, we say so plainly below.

Security by design

This website is deliberately minimal and auditable:

• Static, server-less. It’s a static Jekyll build served by GitHub Pages — no application server, database, or session store collecting visitor data.
• No third-party tracking. No analytics, advertising, social, or A/B-testing scripts; no tracking cookies or fingerprinting.
• No third-party requests on page load. Fonts are self-hosted and all JavaScript is first-party, so simply visiting the site doesn’t call out to Google Fonts, CDNs, analytics, or other third parties that would see your IP address. (The one time data leaves to a third party is when you submit a form — see below.)
• Open source. The full source is public at github.com/CUPIDS-Lab, so anyone can verify these claims.
• Encrypted transport. Served over HTTPS.
• A web form is not a secure channel. Our intake and newsletter forms submit to Formspree, a third-party form processor — your message and whatever you put in it are transmitted to and stored by Formspree under its own privacy policy, then emailed to us. That’s fine for routine, non-sensitive requests; it is explicitly not for confidential sources, privileged material, or anything you need kept off third-party servers. For those, use Secure contact on the Get Involved page.

Built with AI — and accountable for it

We ask others to disclose how their tools work, so we disclose how this site is made. CUPIDS was designed and built with AI assistance — a combination of Claude Design (initial design exploration) and Claude Code (implementation) — and we continue to use AI tools to help maintain and extend it.

• People are accountable, not the model. AI assists; named members of the lab review, approve, and answer for everything that ships. Nothing reaches the site without human review.
• Auditable in the open. Because the full source is public, our use of AI is verifiable in the commit history rather than merely asserted — every change is attributable and reviewable, including the ones AI helped write.
• No reader data feeds any model. The static build above collects nothing about visitors, and we send nothing about you to any AI service. AI helps us write the site; it is never used to watch the people who read it.

Our disclosure infrastructure stays off CU IT

CUPIDS is based at the University of Colorado, but the infrastructure we use to receive sensitive material and to communicate with sources is kept deliberately separate from University of Colorado IT systems.

• Secure channels — Signal, PGP-encrypted email on non-university keys, and our secure file drop — do not run on CU-operated servers, accounts, or networks.
• This site is hosted independently (GitHub Pages), not on university infrastructure.
• We avoid routing confidential source communications through university email, storage, or logging, which can be subject to institutional retention, monitoring, or legal process outside our control.

The goal is simple: people who help us hold power to account should not have their identity or materials exposed by the systems of any single institution — including our own. (Some routine, non-sensitive contact still uses university addresses; for anything sensitive, use the dedicated secure channels.)

Protecting sources & whistleblowers

• Data minimization. We ask for only what we need and encourage you to share only what’s necessary. Don’t send identifying details you don’t have to.
• Use protected channels. For sensitive material, reach us via Signal or our secure file drop rather than email or web forms. Consider using Tor and a device or account not tied to your identity or employer.
• We don’t unmask sources. We will not voluntarily disclose the identity of a source or the existence of a confidential contact, and we will resist improper or overbroad demands to the fullest extent we are able.
• Retention. Sensitive materials are stored encrypted, access is limited to the people doing the work, and we delete what we no longer need.

What we can — and can’t — promise

We’re a public-interest research lab, not a law firm, and this infrastructure is actively being built out. We’d rather be honest about the limits:

• No system is perfectly secure and no promise of anonymity is absolute — your own operational security matters.
• We cannot offer legal privilege, and we can’t guarantee protection against every legal or technical adversary.
• Where a channel above says “on request” or “to be configured,” it is still being stood up. Ask us for current, verified contact details before sending anything sensitive.

Responsible AI in the tools we build

Preserving and analyzing public records increasingly means using AI — to read documents at scale, classify and cluster records, and search large corpora. We hold the tools we build, evaluate, and recommend to the same public-interest standard as the rest of our work, and we put the limits up front rather than burying them.

• Models are instruments that need calibration. We sample for accuracy, measure and report error rates, and document intended use and limits with Model Cards, guided by the NIST AI Risk Management Framework. See our evaluated tooling for how we test a tool before we rely on — or recommend — it.
• Humans stay in the loop. AI output is a lead, not a finding. People verify before anything is published or relied on — especially in journalism, where a confident, wrong machine answer is a liability.
• Sensitive material stays off third-party AI. We do not paste confidential source material, privileged documents, or anything sent through a secure channel into external AI services that would transmit, store, or train on it.
• No AI surveillance of sources. We do not use AI to de-anonymize, re-identify, or profile sources, whistleblowers, or the people who appear in the records we handle.

Responsible disclosure

Found a security problem with this site or our infrastructure? Please tell us privately first via a secure channel on the Get Involved page, and give us a reasonable window to fix it before disclosing publicly. We’re grateful for the help.

This is an initial draft policy and will evolve as the lab’s infrastructure matures. Last reviewed: 2026-06.