Shared-GPU side channel can infer an agent's approval policy
Start here
In the PolicySniff preprint submitted to arXiv on November 12, 2026, Weimin Fu and other authors report that an observer sharing GPU execution with an agent can infer the approval rules it is using. The experiments were limited to the eager sparse mixture-of-experts runtime using Qwen3-30B-A3B-Instruct-2507, conducted on a single A100 GPU, with the observer and agent sharing one MIG instance, and with the candidate rules and the specific task known; observation also caused noticeable slowdown. In the main evaluation on tasks without prior authorization, the balanced accuracy for identifying the current setting was 94.17%, and 88.82% on tasks where the two settings produced the same decision. The authors emphasize that this is an experimental result, does not constitute a general attack against agents in deployment, and did not achieve arbitrary prompt recovery or unauthorized operations.
Generated by AI from reports · Updated 4 days ago
Attention and developments
Attention, hourlyA report
Now 37, 0% in 24 hours·Peak 40 (Nov 16 11:00)·1 report in this range, 1 earlier
Progress timeline
- CASRAIPolicySniff preprint: agent approval policies leak on shared GPUs
In the PolicySniff preprint submitted to arXiv on November 12, 2026, Weimin Fu and other authors report that shared GPU execution can infer the approval rules an AI agent is using, with a balanced accuracy of 94.17% in the main evaluation on tasks without prior authorization for identifying the current setting, and 88.82% on tasks where the two settings produce the same decision. The experiments were limited to an eager sparse mixture-of-experts runtime using Qwen3-30B-A3B-Instruct-2507, the coverage study was conducted on a single A100 GPU, the observer shared a MIG instance with the agent, and the observer knew the candidate rules and the specific task; observation also caused a noticeable slowdown. The article distinguishes the problem into three independent evaluation questions—behavioral compliance, configuration confidentiality, and authorization enforcement—and emphasizes that these are the authors' experimental results, do not constitute a general attack against deployed agents, and did not achieve arbitrary prompt recovery or unauthorized operations; arXiv moderation is not peer review.
- arXiv: 其他 AI 安全论文PolicySniff: inferring an agent's policy state through shared GPU execution
Researchers propose the PolicySniff attack, which infers an agent's approval policy state by executing on a shared GPU; on an eager sparse MoE runtime with selected experts running independently, a frozen and blinded study identified the deployed setting with 94.17% balanced accuracy. LLM agents set approval policies via the system prompt, and operators keep them secret by restricting access to the prompt and responses, while existing audits only check whether the agent complies with the setting, not whether an adjacent process can infer the setting. The attacker first models the two candidate settings separately, then measures only the timing of its own GPU kernels while the agent processes tasks; because the policy prefix is already cached, what the attacker observes is the same task-suffix processing under both settings. In tests covering all authorization conditions, accuracy was 88.82% on tasks where the two settings make the same decision and 91.98% on tasks where the decisions differ; hardware replay of expert routing shows that timing density and layer-to-bucket alignment carry policy information.
Related discussion
- Older than 72 hours · Not included in the 72-hour figures
- X@FSFGPolicySniff: Inferring Agent Policy States through Shared-GPU Execution Weimin Fu, Minghao Shao, Zeng Wang, Qiying Li,…(opens the original post in a new tab)
Attention trend
New sources per day
- Nov 13: 1 new sources (1 outlets, 0 accounts)
- Nov 14: 1 new sources (0 outlets, 1 accounts)
- Nov 15: 0 new sources (0 outlets, 0 accounts)
- Nov 16: 1 new sources (1 outlets, 0 accounts)
- Nov 17: 0 new sources (0 outlets, 0 accounts)
- Nov 18: 0 new sources (0 outlets, 0 accounts)
- Nov 19: 0 new sources (0 outlets, 0 accounts)
- Nov 20: 0 new sources (0 outlets, 0 accounts)