Third-party witness networks are an effective defense against split-view attacks on transparency logs, but they require coordinating multiple independent organizations — incurring reliability, legal, and operational cross-org constraints and risks.
This talk presents LogCabin, an open-source, split-view resistant ledger service built on a Nimble-inspired protocol (Nimble: OSDI ‘23). When deployed as a witness, LogCabin enables a new scenario: a single operator can offer a scalable t-log where the operator itself is technically — not just practically — incapable of creating undetected split views. As a structurally novel type of witness, it contributes to enriching existing witness networks.
LogCabin uses a quorum of endorsers, each running as a Trusted Execution Environment (TEE) application on Oak Restricted Kernel (Project Oak), to collectively maintain append-only ledgers. Each append requires a majority of endorsers to independently sign the new ledger state. Endorser signing keys are generated ephemerally within TEE-protected volatile memory and bound to the binary through a DICE attestation chain, so relying parties can use remote attestation to cryptographically verify that signatures were produced by a specific, measured build. The endorser’s private key never leaves the TEE, and the protocol core is a Rust no_std, zero-I/O pure state machine — compact and auditable.
While LogCabin can be used as a powerful standalone ledger, its impact magnifies when deployed as a witness to a high-throughput Merkle-tree-based transparency log, with the overall system still speaking standard, widely adopted t-log protocols. This approach also shifts the bar for collusion-based split views: rather than requiring the log operator to collude with third-party witnesses, collusion is required with the TEE hardware manufacturer.
LogCabin’s strong integrity guarantees come with real operational challenges. The protocol prioritizes correctness over availability: if conditions for safe operation are not met, LogCabin will fail rather than drop its integrity guarantees. If a majority of endorsers in a cohort is permanently lost, the instance itself is irrecoverably lost. Achieving production-grade, continuous, reliable operation — including TEE lifecycle management, cohort reconfiguration under load, and resilience to infrastructure failures — is a hard problem. In this talk, we explore proposals to achieve high levels of reliability with LogCabin and share early results.
At present, the LogCabin core endorser TEE application is complete and open-sourced, while the coordination, interface, and storage layers are under construction.
Speaker
Ernesto is a software engineer at Google DeepMind working on Security and Privacy, where he designs systems to establish cryptographic trust and transparency for server and AI workloads. His work spans confidential computing, Trusted Execution Environments (TEEs), and verifiable systems, with his current focus centering on the emerging challenges of agentic security and privacy.
