TaskBound Frames the 'Hijacked Authorized Agent' Problem for HPC Agents Running as You

arXiv 2607.18485·low signalXBlueskyLinkedInCopy link

LLM agents doing routine HPC work — monitoring Slurm jobs, diagnosing failed builds, inspecting simulation output — typically run under the user's own credentials and inherit full access to files and the scheduler. The paper names the resulting gap the hijacked authorized agent problem: adversarial instructions planted in a log, tool description, shared file, or peer-agent message can redirect the agent well beyond its assigned task while every resulting command remains authenticated and permitted for that account, so account-level controls register nothing wrong. The authors propose TaskBound as a benchmark for empirical evaluation, though this is primarily threat modeling rather than measured results.