Workshop: AgenticAI4HPC’26: The First International Workshop on Agentic AI for HPC
Authors: Emily Herron, Jack Hutchins, Amir Sadovnik, Jesse Hines, Vanessa Lama, Junqi Yin, and Feiyi Wang (Oak Ridge National Laboratory)
Abstract: For LLM agents at DOE facilities, prompt injection and unsupported citations share one shape: a value reaches a consequential action, a sink, from a forbidden origin. PALISADE, a sidecar gating six boundaries of the VISTA platform, labels values by origin outside the model, propagates labels by text matching, and admits a sink value only if its label permits. SIEGE, our 205-attack benchmark, enforces HPC job submission as a security boundary and scores detector evasion separately from sink access. The deterministic tier stops 19 of 32 sink-capable attacks and, with egress blocking, all 5 forged citations, at 0.28 ms/turn and 3.3% benign blocking. Adaptive attacks (PAIR, TAP) raise credited sink admissions to means of 28–31 of 32, but about 14 still carry the attack, none bearing an untrusted label at the check. Closing this propagation gap requires tracking dataflow or keeping untrusted values out of the planner.
Back to AgenticAI4HPC’26: The First International Workshop on Agentic AI for HPC Archive Listing
Back to Full Workshop Archive Listing