sectum-ai

Sectum AI is an open-source platform designed to verify and provide evidence of tenant isolation in multi-tenant AI systems. It addresses the challenge of ensuring that data from one customer (tenant) cannot be accessed by another within environments such as vector databases, retrieval-augmented generation (RAG) pipelines, agent frameworks, semantic caches, fine-tuned models, and managed compute providers. The platform generates tamper-evident attestations, including a PDF and a machine-readable evidence pack, which can be independently verified by auditors, security teams, and data protection officers without requiring trust in Sectum AI itself.

The tool operates by seeding synthetic tenants with cryptographic canary markers and a hashed ground-truth manifest, enabling deterministic and reproducible verification with zero false positives. Sectum AI tests across thirteen surfaces, including vector databases, RAG pipelines, semantic and key-value caches, agent memory, tool calls, fine-tunes or adapters, evaluation sets, backups, search indexes, tracing pipelines, prompt/completion logs, and APIs. It supports live adapters for common backend systems. The platform is equipped to detect eleven classes of cross-tenant attacks, such as direct tenant-boundary fetches, organic entity-bleed in RAG, semantic-cache contamination, timing side channels in key-value caches, embedding inversion, confused-deputy bugs, persistent memory contamination, LoRA cross-tenant influence, benign extraction, RAG poisoning, and GDPR Article 17 erasure verification.

Sectum AI outputs are canonicalized, hashed, RFC 3161 timestamped, logged with Sigstore Rekor, and wrapped in an in-toto attestation envelope. The resulting evidence can be rendered as an auditor-ready PDF, and the verification process can be independently run using the sectum-ai verify tool, which does not require a Sectum AI installation. The platform is particularly suited for roles such as CISOs, sales engineers, compliance teams, platform engineers, application security professionals, legal teams, and DPOs, supporting use cases like vendor security questionnaires, SOC 2 audit evidence, pre-launch verification, CI regression baselines, GDPR erasure response, and compliance with the EU AI Act.

0 license, with its marker substrate, attack catalog, adapters, evidence chain, and verification tool available as open source. The platform is designed to be integrated into audit and compliance workflows, enabling partners such as audit firms to white-label its evidence packs for standards like SOC 2, ISO 27001, and GDPR. Sectum AI distinguishes itself by enabling real tenant provisioning to measure cross-tenant leakage and by producing independently verifiable evidence, addressing a gap not filled by LLM red-team frameworks, runtime guardrails, governance platforms, or data security tools.