Show HN: Distributed, deterministically secure agentic swarm runtime. Enclawed (zenodo.org)

A new project named Enclawed introduces a deterministic security runtime designed for large language model (LLM) agents operating on high-performance computing (HPC) systems. Unlike current systems that lack proper job supervision and are vulnerable to compromised control hosts, Enclawed enhances security by implementing a consensus-governed, hardware-rooted supervision mechanism. The system utilizes an "Enclaweder," which acts as a hardware root of trust to prevent catastrophic failures, ensuring that simulated jobs are securely launched and monitored. Each node within this swarm system can only execute tasks that have been authorized by a clear policy, effectively mitigating the risk of malicious interference. This development is particularly significant for the AI/ML community, as it addresses critical security concerns linked to LLM agents and high-stakes simulations, where inaccuracies can lead to severe consequences. Enclawed operates without requiring special administrative privileges and integrates seamlessly into existing HPC environments via Secure Shell, adding minimal overhead. The project has successfully demonstrated its robustness against multiple live attacks, rejecting all unauthorized actions while allowing legitimate ones, thus establishing a new benchmark for security standards in AI agent deployments. As an open-source initiative, both Enclawed and the accompanying HPC Agent Security Benchmark promise to foster greater innovation and collaborative improvement in securing AI systems.