Loops HousePrivateTriage

NEAR HACKATHON
Links
Team
1 member- MIOwner
Mickey Joe
Overview
PrivateTriage — attest before you disclose.
Unpublished vulnerability reports can contain sensitive exploit details. Sending them to hosted AI normally requires trusting the provider's environment. PrivateTriage is an open-source vulnerability-triage prototype built on NEAR AI Cloud that verifies the confidential-computing environment before sharing a report.
For every request it generates a fresh challenge tied to the exact request hash, verifies Intel TDX attestation for the gateway and model, binds TLS to the attested gateway, checks NVIDIA GPU confidential-computing evidence, and enforces an explicit measured-software policy. Only after those preflight checks pass does it transmit the report over the same pinned connection. Afterwards, it requires a model-TEE signature covering the exact request and response hashes. Missing or invalid evidence never receives a VERIFIED label.
The AI produces structured triage with severity, CVSS vector, impact, prerequisites, remediation and questions for the reporter. PrivateTriage independently recomputes CVSS scores, flags suspected prompt injection, and clearly separates model recommendations from cryptographically verified execution.
Real NEAR AI Cloud inference on three fictional reports produced archived evidence that anyone can re-verify offline without an API key. The application displays five verification properties and 47 checks, plus 26 tested tampering scenarios (21 FAILED, five UNVERIFIED), supported by 113 automated tests. The public guided demo replays these genuine historical runs; live inference remains local.
This is a second independent NEAR Hackathon entry, distinct from Pursewarden, for Privacy & Private AI. Research prototype, not audited. Attestation does not establish AI correctness, provider non-retention or end-to-end encryption to the model.
Guided demo: https://mickeyjoes.github.io/near-private-triage/