A colocated observer classifies a target agent's concealed approval rules from GPU timing while receiving neither its prompt nor its model responses. The attacker benchmarks policy variants, holds the victim's public task suffix constant and infers the active internal approval condition from interference with scheduled observer kernels in a shared…
Inspect the ClaimsCase · DiggingBeagle record
PolicySniff reads agent approval-policy state through shared-GPU timing signals
A colocated observer classifies a target agent's concealed approval rules from GPU timing while receiving neither its prompt nor its model responses.
Calibrated A100 MIG and quantized gpt-oss-120b experiments using known candidate policies and task families; not a demonstrated cross-provider breach.
Assessment profile
Assessment method- Setting
- research demonstration
- Exploitation
- demonstrated
- Evidence
- Grade C
- Remediation
- none known
- AI role
- AGAINST AI
Basis and provenance
A frozen classifier identifies policy states above chance in the test environment, with intrusive probe slowdown and no demonstrated arbitrary secret extraction. Classified as research_demonstration with exploit status demonstrated. Evidence grade C reflects the original paper or maintainer advisory and its stated limitations, not an independent audit of all deployments. Comparable realized harm is not established, so impact stays unrated; remediation status none_known is limited to the cited experiment or affected software version. New source-specific boundary: The same probing that provides a stronger policy fingerprint can impose measurable service slowdown, making inference accuracy and attack observability coupled rather than independent.
Assessed Nov 15, 2026 using diggingbeagle.assessment/1.
Evidence basis: CLM-R245-POLICYSNIFF-GPU-TIMING-01 · CLM-R245-POLICYSNIFF-GPU-TIMING-02 · CLM-R246-POLICYSNIFF-GPU-TIMING-BOUNDARY · CLM-ENR250-12 · PolicySniff reads agent approval-policy state through shared-GPU timing signals (research paper)
At a glance
The attacker already knows the small hypothesis set and shares controlled hardware; this is not arbitrary system-prompt reconstruction or an unobtrusive production side channel. No evidence here establishes the number of production victims or a broader failure rate outside the cited evaluation.
Limits and uncertaintyFull account
A colocated observer classifies a target agent's concealed approval rules from GPU timing while receiving neither its prompt nor its model responses.
The attacker benchmarks policy variants, holds the victim's public task suffix constant and infers the active internal approval condition from interference with scheduled observer kernels in a shared accelerator. Authors report 94.17 percent exact-task balanced policy-state accuracy over 120 clusters and 88.82 percent accuracy even when visible tool actions agree, although continuous probing slows victim execution 4.71 times.
The attacker already knows the small hypothesis set and shares controlled hardware; this is not arbitrary system-prompt reconstruction or an unobtrusive production side channel.
The attacker does not read the policy text. It colocates an observer workload with a victim inference process on the same GPU partition, calibrates timings for known possible approval policies and then uses differences in probe delay to classify the policy state. The evaluated policy prefix is cached and the visible user task held constant, making the signal different from an ordinary answer or tool-action comparison. This is discrimination among predefined policy options, not arbitrary prompt extraction.
The observer also perturbs the system it measures. Continuous probing produces a large latency penalty, and a less aggressive cadence loses accuracy. The study requires known candidate implementations, exact-task calibration and shared hardware conditions. Protecting an agent from unauthorized calls therefore does not itself guarantee confidentiality of its policy, but the paper does not establish cross-tenant leakage on an independently isolated cloud service.
Mechanism and trust boundary
- 01
Entry or starting condition
An attacker can run calibrated GPU probes in the same A100 MIG environment as a victim agent inference workload.
Boundary: attacker or research precondition
- 02
Crossing the relevant boundary
Shared GPU execution timing carries a signal correlated with one of several known hidden authorization conditions.
Boundary: mechanism and authorization boundary
- 03
Observed or reported result
A frozen classifier identifies policy states above chance in the test environment, with intrusive probe slowdown and no demonstrated arbitrary secret extraction.
Boundary: measured outcome and evidentiary limit
Claims & evidence
CLM-ENR250-12In the paper's calibrated A100 MIG setup, rapid 1 ms probes produced about 93.75 percent policy-classification accuracy in the measured configuration. The attack requires a co-resident timing observer, a restricted set of known policy candidates and frequent measurements; those prerequisites differ from recovering an arbitrary private system prompt.supported
Basis: reported finding
- supportsPolicySniff reads agent approval-policy state through shared-GPU timing signals (research paper)primary disclosure
Evaluation of probe interval, 1 ms cadence, policy accuracy and GPU co-residency assumptions
CLM-R245-POLICYSNIFF-GPU-TIMING-01Authors report 94.17 percent exact-task balanced policy-state accuracy over 120 clusters and 88.82 percent accuracy even when visible tool actions agree, although continuous probing slows victim execution 4.71 times.supported
Basis: reported finding
- supportsPolicySniff reads agent approval-policy state through shared-GPU timing signals (research paper)primary disclosure
Results, exact-task and action-agreement policy classification tables
CLM-R245-POLICYSNIFF-GPU-TIMING-02The attacker benchmarks policy variants, holds the victim's public task suffix constant and infers the active internal approval condition from interference with scheduled observer kernels in a shared accelerator.supported
Basis: reported finding
- supportsPolicySniff reads agent approval-policy state through shared-GPU timing signals (research paper)primary disclosure
Threat model and shared-GPU observer methodology
CLM-R246-POLICYSNIFF-GPU-TIMING-BOUNDARYThe PolicySniff paper reports that continuous probing increases median victim latency by 4.71 times, whereas adding a 5 ms gap reduces slowdown to 1.52 times while reducing policy inference accuracy to 58.33 percent in the evaluated cadence experiment.supported
Basis: reported finding
- supportsPolicySniff reads agent approval-policy state through shared-GPU timing signals (research paper)primary disclosure
Section VI-D, Table IV Probe cadence and victim cost
Unknowns and contradictions
- The attacker already knows the small hypothesis set and shares controlled hardware; this is not arbitrary system-prompt reconstruction or an unobtrusive production side channel.
- No evidence here establishes the number of production victims or a broader failure rate outside the cited evaluation.
Sources and citation
Material revision history
- Nov 15, 2026 · Canonical change recorded · new in release · revision 251