arXiv:2611.16158 · v1

PolicySniff: Inferring Agent Policy States through Shared-GPU Execution

首次提交 2026-11-12 09:42

版本更新 2026-11-12 09:42

雷达观测 2026-11-13 14:16

摘要

LLM agents apply approval settings, such as whether to ask before a consequential tool call, through instructions in their system prompt, and operators keep these settings confidential by restricting access to prompts and responses. Audits check whether an agent obeys its setting, not whether a neighboring process can infer which setting is deployed. PolicySniff infers an agent's approval-policy state through shared-GPU execution. A co-resident attacker that knows the candidate settings and the task first profiles both settings, then times only its own GPU kernels while the agent processes the task. The policy prefix is cached before capture, so the attacker observes processing of a task suffix that is identical under both settings. On an eager sparse mixture-of-experts runtime, where each selected expert runs as separate work, a frozen and blinded study identified the deployed setting at 94.17% balanced accuracy on its primary estimand. Across all authorization conditions, accuracy was 88.82% on tasks for which both settings produce the same decision and 91.98% on tasks where they differ. Hardware replays of expert routing show that its temporal density and layer-and-bucket alignment carry the policy information. Behavioral compliance does not imply configuration confidentiality: deployments that keep their approval setting private should measure its distinguishability on the serving runtime and enforce authorization outside the model.