Architecture
Security
PromptSlayer is built safe-by-default. This page describes how the architecture enforces authorization, isolation, and data minimization.
No anonymous scanning
There is no anonymous scan mode. The live audit form requires explicit target declaration and an authorization checkbox before any request is sent.
Demo mode isolation
The public demo never issues outbound requests to real targets. Phase progression, evaluator verdicts, and event feeds are simulated in the browser, so reviewers can explore the tool without capability risk.
Authorization first
Every live audit requires the operator to confirm in-app that they are authorized to test the configured target. Without this acknowledgment, the Start Audit button stays disabled.
Data minimization
The frontend holds no telemetry infrastructure. Audit history stays in localStorage on the operator's device, and audit results are transmitted only between the configured backend and the operator's browser.
Result handling
Audit results are stored by the backend service that the operator configures via VITE_API_URL. PromptSlayer does not provide backend hosting and does not retain results on its own infrastructure.
Credentials
API keys entered into the audit form are held in component state for the duration of the session. They are not written to localStorage or transmitted to third parties.
Disclosure practice
Vulnerabilities discovered against third-party models are reported to the vendor first through their published security channel, with a reproduction summary and suggested mitigation. Public write-ups describe the technique class and its impact, never a copy-paste payload, and only after remediation or an agreed disclosure deadline.
Open by design
The codebase is auditable. Defensive use and coordinated disclosure are explicit values, not marketing copy.