PolicySniff: Inferring Agent Policy States through Shared-GPU Execution
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.
Nadia: We're looking at the paper: "PolicySniff: Inferring Agent Policy States through Shared-GPU Execution".
Elias: Here's the gist of the paper. LLM agents apply approval settings through instructions in their system prompt, and PolicySniff infers an agent’s approval-policy state through shared-GPU execution 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, a frozen and blinded study identified the deployed setting at 94.17% balanced accuracy on its primary estimand 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 Confidentiality Target The paper defines a confidentiality target for agent deployments as the identity of a behaviorally validated approval setting, defined against a profiled co-resident attacker that never queries the service The reconnaissance target is the identity of the active rule, and its recovery reveals an authorization-relevant configuration even if the attacker cannot reconstruct the prompt or observe an action that distinguishes the candidates Native Attack Mechanism The native attack follows policy state through GPU execution to timing and estimate Sˆ The mechanism follows from how cached context affects sparse inference, where each task token attends to the cached policy prefix, and the resulting hidden states influence which experts run in a sparse mixture-of-experts (MoE) model The attacker profiles each candidate policy on the shared GPU for a known task, then assigns the target trace to the closer reference Evidence Structure and Methodology The paper tests three paths from policy to observation: behavioral validation follows πS to A, the native attack follows πS through E to X and Sˆ, and replay experiments follow πS through R and Z to a physical replay The native attack observes ordinary eager execution of the unmodified model; it receives no routing summary and needs no cooperating sender Mechanism Evidence Routing replays show that temporal burst density and layer-and-bucket alignment of expert routing carry the policy, while constant-slot padding suppresses the evaluated decoders at a measured activity cost The temporal intervention fixes each request’s byte budget, and the spatial intervention fixes per-layer work between paired allocations Security Implications Policy identification and action redirection have separate success criteria: the first recovers a configuration, and the second induces an unauthorized action The results establish suppression of the evaluated classifier when authorization is represented outside the measured inference path under this fixed interface PolicySniff shows that physical execution can disclose an agent’s approval setting while its prompt and responses stay hidden The disclosure reveals which deployments instruct the model to proceed without human confirmation and can inform target selection for subsequent jailbreak or promptinjection attacks The paper establishes that policy-state distinguishability next to serving cost as an evaluation criterion The comparison suppresses the evaluated classifier when authorization is represented outside the measured inference path under this fixed interface Measurement and Mitigation Continuous probing slowed the victim by a median factor of 4.71× and probing with a 1 ms idle interval by 2.96× Mitigation can address the observer’s access, the execution feature that carries the policy, or the placement of the policy itself Runtime developers can evaluate execution strategies by policy-state distinguishability as well as serving cost The comparison moves the rule to a broker with trusted operation and authorization metadata, suppressing the evaluated timing classifier Policy Placement Comparison The guarded arm instructs the model to emit a fixed execution proposal and applies the rule after inference with a trusted broker The baseline balanced accuracy is 95.83% at 100.
Paper summary: Nadia: So to wrap up, this paper PolicySniff: Inferring Agent Policy States through Shared-GPU Execution shows that agent approval settings can be inferred by a co-resident attacker by timing GPU kernels during task processing, even when the agent tries to keep those settings private through system instructions.
Elias: The authors are focused on the mechanism of how cached policy context affects expert selection during task processing, and they found specific routing patterns, like temporal burst density and layer and bucket alignment, carry this policy information.
Priya: It’s important to remember that behavioral compliance doesn't equate to configuration confidentiality; deployments need to measure the setting's distinguishability on the serving runtime while enforcing authorization outside that model inference path.
Nadia: The paper suggests that physical execution can disclose an agent’s approval setting even if its prompt and responses stay hidden, revealing which deployments instruct the model to proceed without human confirmation, which informs target selection for subsequent attacks.
Elias: Ultimately, the work establishes that policy-state distinguishability next to serving cost is a key evaluation criterion for security against these kinds of inference attacks. The authors conclude that runtime developers can evaluate execution strategies using this metric.
Conclusion: Nadia: So, to put it simply, this paper shows that you can figure out exactly which approval setting an agent is using just by watching how it runs on a shared GPU while it's doing a task.
Elias: It’s about inference through timing and execution patterns on the hardware itself. The authors are looking at how caching a policy prefix affects which parts of the model get used in a sparse mixture-of-experts setup.
Priya: What this means is that behavior alone doesn't tell you what's configured; you have to measure how distinguishable those settings are when they’re actually running on the serving system.
Nadia: Exactly. The paper explores how temporal density and layer and bucket alignment in the expert routing carry that policy information, which is a way of seeing it happening physically on the GPU.
Elias: They show that you can recover authorization-relevant configuration just by looking at those routing replays, even if you can't reconstruct what the agent was actually prompted with.
Priya: It really highlights the gap between what an agent claims to be doing and what's actually happening in the execution path.
Nadia: If this holds up, it means we can use physical execution traces to identify risky deployments that are bypassing human confirmation on important decisions.
Elias: That kind of insight could inform how we design defenses against prompt injection or jailbreak attempts by understanding which settings are active.
Weimin Fu, Minghao Shao, Zeng Wang, Qiying Li, Yuntao Liu, Johann Knechtel, Ozgur Sinanoglu, Muhammad Shafique, Ramesh Karri, §iaolong Guo
Lehigh University · –New York University Department of Computer Science and Engineering New York University Abu Dhabi
cs.CR
Submitted: 2026-11-12
Updated: 2026-11-12
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 99/100
The gist: LLM agents apply approval settings through instructions in their system prompt, and PolicySniff infers an agent’s approval-policy state through shared-GPU execution The policy prefix is cached…
Key concepts
- Confidentiality Target
- This is the specific identity of a behaviorally validated approval setting that the paper aims to protect. It's defined against an attacker who never directly queries the service, making it a precise target for policy inference.
- Native Attack Mechanism
- The attack follows the agent's policy state through GPU execution to timing and estimate Sˆ. It exploits how cached context affects sparse inference in mixture-of-experts models, where task tokens attend to a cached prefix, influencing which experts are activated.
- Mechanism Evidence
- Routing replays reveal that temporal burst density and layer-and-bucket alignment of expert routing carry policy information. Conversely, constant-slot padding suppresses the evaluated decoders at a measured activity cost, showing how execution patterns reveal configuration details.
- Policy Placement Comparison
- This compares two deployment strategies: one where the guarded arm emits a fixed execution proposal and applies the rule after inference via a trusted broker, versus a baseline. This comparison helps developers evaluate execution strategies based on policy-state distinguishability and serving cost.
Terminology
Summary
LLM agents apply approval settings through instructions in their system prompt, and PolicySniff infers an agent’s approval-policy state through shared-GPU execution 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, a frozen and blinded study identified the deployed setting at 94. p. 017% balanced accuracy on its primary estimand 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
Confidentiality Target
The paper defines a confidentiality target for agent deployments as the identity of a behaviorally validated approval setting, defined against a profiled co-resident attacker that never queries the service The reconnaissance target is the identity of the active rule, and its recovery reveals an authorization-relevant configuration even if the attacker cannot reconstruct the prompt or observe an action that distinguishes the candidates
Native Attack Mechanism
The native attack follows policy state through GPU execution to timing and estimate Sˆ The mechanism follows from how cached context affects sparse inference, where each task token attends to the cached policy prefix, and the resulting hidden states influence which experts run in a sparse mixture-of-experts (MoE) model The attacker profiles each candidate policy on the shared GPU for a known task, then assigns the target trace to the closer reference
Evidence Structure and Methodology
The paper tests three paths from policy to observation: behavioral validation follows πS to A, the native attack follows πS through E to X and Sˆ, and replay experiments follow πS through R and Z to a physical replay The native attack observes ordinary eager execution of the unmodified model; it receives no routing summary and needs no cooperating sender
Mechanism Evidence
Routing replays show that temporal burst density and layer-and-bucket alignment of expert routing carry the policy, while constant-slot padding suppresses the evaluated decoders at a measured activity cost The temporal intervention fixes each request’s byte budget, and the spatial intervention fixes per-layer work between paired allocations
Security Implications
Policy identification and action redirection have separate success criteria: the first recovers a configuration, and the second induces an unauthorized action The results establish suppression of the evaluated classifier when authorization is represented outside the measured inference path under this fixed interface PolicySniff shows that physical execution can disclose an agent’s approval setting while its prompt and responses stay hidden The disclosure reveals which deployments instruct the model to proceed without human confirmation and can inform target selection for subsequent jailbreak or promptinjection attacks The paper establishes that policy-state distinguishability next to serving cost as an evaluation criterion The comparison suppresses the evaluated classifier when authorization is represented outside the measured inference path under this fixed interface
Measurement and Mitigation
Continuous probing slowed the victim by a median factor of 4.71× and probing with a 1 ms idle interval by 2. p. 696× Mitigation can address the observer’s access, the execution feature that carries the policy, or the placement of the policy itself Runtime developers can evaluate execution strategies by policy-state distinguishability as well as serving cost The comparison moves the rule to a broker with trusted operation and authorization metadata, suppressing the evaluated timing classifier
Policy Placement Comparison
The guarded arm instructs the model to emit a fixed execution proposal and applies the rule after inference with a trusted broker The baseline balanced accuracy is 95.83% at 100. p. 1
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Confidentiality target for agent deployments: Define a security-relevant label as
the identity of a behaviorally validated approval setting
to identify which deployment instructs the model to proceed without human confirmation. This allows operators to assesswhich execution observations reveal a rule whose choice they treat as private.
-
Response-free native attack on eager sparse MoE runtime: Implement an attacker that knows candidate settings and task first profiles both, then
times only its own GPU kernels while the agent processes the task,
identifying the deployed setting at 94.17% balanced accuracy on its primary estimand. -
Mechanism evidence from routing replays: Analyze temporal density and layer-and-bucket alignment of expert routing to identify policy information, as
temporal burst density and layer-and-bucket alignment of expert routing carry the policy information.
-
Policy placement comparison under a fixed action interface: Move the approval rule to a trusted broker with trusted operation metadata to
suppress the evaluated timing classifier while preserving the prescribed decisions.
Sources
- Has My System Prompt Been Used? Large Language Model Prompt Membership Inference
- Mixtral of Experts
- Qwen3 Technical Report
- Stealing User Prompts from Mixture of Experts
- Misrouter: Exploiting Routing Mechanisms for Input-Only Attacks on Mixture-of-Experts LLMs
- Remote Timing Attacks on Efficient Language Model Inference
- NVBleed: Covert and Side-Channel Attacks on NVIDIA Multi-GPU Interconnect
- PAGE404: Breaking GPU Isolation Through Hardware-assisted Page Migration
Related papers
- From Strategy Search to Policy Learning: MCTS-Guided Adversarial Co-Evolution for LLM Safety Alignment
- Privacy Leakage from Recurrent State in LLM Inference
- Incrementally Verifiable Oracle Aided Computation
- Privilege and confidentiality in generative AI workflows
- BadMoE: Backdooring Mixture-of-Experts LLMs via Routing-Aware Triggers and Task-Underutilized Experts
- CoVault: Secure, Scalable Analytics of Personal Data