noirebox
Tamper-evident journal with Ed25519 signatures for AI agent auditability, RFC 3161 timestamp anchoring and third-party verifiable exports.
Ficha actualizada 22 sept 2026, 18:54 · upstream 22 sept 2026, 17:32 · release sin release detectada
Qué es noirebox
Descripción editorial construida a partir del repositorio, su documentación y sus señales de mantenimiento.
NoireBox is a tamper-evident flight data recorder designed specifically for AI agents. It records every decision made by an agent in a hash-chained journal signed with Ed25519, anchored to RFC 3161 timestamps via a Timestamp Authority (TSA). This provides verifiable proof of what was produced, when it occurred, and on which inputs — without requiring trust in the operator. The system flags poisoned transcripts before they reach the agent and exports an attestation that anyone can verify offline using only the journal itself.
NoireBox is designed for European markets where GDPR and the EU AI Act require demonstrable accountability. It answers three core obligations: proving what an AI produced, on which data, and when; enabling data subject rights through signed export of meeting content; and maintaining automatic event logging required by the EU AI Act for risk systems.
The product is built in France with a focus on sovereignty — self-hosted deployment without cloud dependencies, Ed25519 keys stored locally, and no telemetry. It applies to any agent or service: from local Ollama models running on a laptop to enterprise-grade deployments across multiple fleets.
- Audit an LLM's output for compliance with GDPR requirements (art. 5(2) accountability)
- Verify that an AI agent did not generate poisoned or malicious content before deployment
- Provide verifiable proof of meeting notes to a client or auditor without trusting the operator
- Enable data subject rights by exporting signed records of all interactions with a specific user
- Automatically log and attest every decision made by a risk-based AI system under EU AI Act obligations
- Maintain audit trails for software agents that make business decisions in French or English environments
- Provides tamper-evident proof of an agent's output with Ed25519 signatures and RFC 3161 timestamp anchoring, verified offline
- Supports both languages (French and English) through a single pluggable guardrail engine trained on ~5,400 examples per language (~250 KB each)
- Self-hosted deployment with no cloud dependency, using Ed25519 keys stored locally to ensure data sovereignty
- The guardrail is designed as an optional plugin that can be swapped in/out without affecting the core journal functionality
- NoireBox does not provide a full compliance certification; it only provides tamper-evident proof of integrity for specific use cases (GDPR art. 5(2), EU AI Act art. 12)
- The bundled guardrail is a heuristic-based regex + ML model that may miss edge cases or adversarial inputs not covered by the training data
- NoireBox does not provide full auditability for all data types — only structured outputs (transcripts, logs) are supported as input to the journal
- The TSA anchor requires an external Timestamp Authority service; self-hosted OpenSSL is available but adds operational complexity in production environments
- The ML guardrail has a known limitation: it does not detect adversarial attacks that specifically target the model's internal weights, only patterns in the output text
Perfil de infraestructura
Estimación inicial derivada del repositorio y de los servicios detectados.
Despliegue detectado
Qué señales reales de instalación hemos encontrado.
Deploy NoireBox on any infrastructure supporting Python and SQLite. For local development, run `make demo` which spins up an API server on port 8768 with a micro-model for testing. For production, deploy as a standalone service using Docker Compose or Kubernetes. The journal stores data in SQLite files that can be backed up to any storage system (S3, GCS, local disk). No telemetry is collected; all keys are self-managed via the CLI `noirebox init`. For multi-tenant deployments with multiple agents, use the fleet anchoring feature (`make demo-fleet`) which aggregates proof chains into a single TSA-signed record covering all journals. Full deployment guide: docs/DEPLOIEMENT.md
Stack y dependencias
Componentes inferidos a partir de manifests, configuración y documentación.
Puede sustituir a
Relaciones detectadas por el análisis del repositorio y su categoría.
Experiencia real de la comunidad
Ayuda a contrastar la estimación automática con instalaciones reales.
Todavía no hay datos comunitarios para este proyecto.