Enclawed adds hardware-rooted supervision for LLM jobs on HPC
Original title: A Deterministic Security Runtime for LLM Agents on MPI and Batch-Scheduled HPC Systems: Consensus-Governed, Hardware-Rooted Supervision of Every Rank
A launch gate denied all 24 planted launches and allowed 108 legitimate ones.
The paper extends Enclawed from a single-user runtime to parallel HPC jobs and agent swarms. A hardware root of trust and a policy-only launch gate are meant to stop a compromised control host from authorizing bad work. Running ranks receive short leases from signed heartbeats and are paused or killed if a lease lapses or tampering is detected. The authors say it needs no administrator rights, uses SSH, and adds about 3 seconds of startup on a 16-rank LAMMPS job.
Unconfirmed: Resistance claims are the authors' lab results, not independent operational confirmation.
- Of 23 surveyed systems, none supervised the running job.
- Twelve live attacks on keyed ranks, including colluding coordinators, failed closed.
- The LAMMPS compute loop was unchanged, including on GPUs.
- Enclawed's core and the HPC Agent Security Benchmark are open source.
From the original article
1 min read at arxiv.orgLarge language model agents launch simulations on high-performance computing (HPC) systems, where one sbatch or mpirun can commit thousands of node-hours and a forged result can cost lives. Of 23 surveyed systems, none supervises the running job, and none holds when the control host is compromised. We fix both.
By Alfredo Metere · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.