Run a local-first healthcare AI prompt-injection challenge, get a survival verdict, inspect baseline agent context, and publish the result as a visual proof dossier, offline proof PDF, generated UI mockup, report, README marker, social card, SARIF file, OpenTelemetry log bundle, JSON, and Markdown.
A useful developer tool needs an outcome people can screenshot, review, and cite from a README.
The default challenge pack returns a result like 8/10 survived, with blocked traps called out as concrete report events.
Every challenge bundle includes badge.svg and README-badge.md so other projects can link directly to their proof packet.
Reports include OpenAI-compatible chat, RAG bot, claims copilot, prior-auth agent, and voice-agent baseline profiles.
Every challenge also exports proof-dossier.html, offline-proof.pdf, and ui-mockup.html so the result is readable without a terminal.
These are representative profiles, not vendor claims. They make the challenge legible to builders who know their architecture.
Generic chat wrapper with refusal copy but limited tool gating.
Retrieval guardrails help on policy questions but not export attempts.
High tool reach makes export and token traps more dangerous.
Human-review routing helps, but policy override prompts still trip risk.
Narrow tooling and refusal language reduce proven exploit paths.
Use --input or --pack to generate your own proof packet.