In brief

  • ML Backdoors: The LADDER attack formulates a multi-objective optimization problem in dual domains to insert robust backdoors into CNNs, achieving a high attack success rate in the low-frequency spectral region.
  • WiFi Side-Channels: LeakyBeam exploits plaintext beamforming feedback in WiFi to accurately detect human occupancy at a distance.
  • Agent Vulnerabilities: Cross-Tool Harvesting and Polluting (XTHP) reveals a novel class of control flow hijacking in LLM agents, successfully compromising real-world agent frameworks.
  • Database Reconstruction: A new attack leverages multidimensional range queries to reconstruct databases efficiently, heavily reducing the mean squared error over previous approaches.

LADDER: Multi-Objective Backdoor Attack via Evolutionary Algorithm

  • This work proposes a multi-objective black-box backdoor attack in dual domains via an evolutionary algorithm, optimizing triggers without requiring prior knowledge about the victim model.

  • The attack demonstrates effectiveness against convolutional neural networks, achieving an average Attack Success Rate (ASR) of 90.23% when utilizing a low-frequency trigger design.

  • The low-frequency triggers outperform five other backdoor attacks and three LADDER variants in the low-frequency spectral region by 50.09% on average.

  • Dazhuang Liu, Yanqi Qiao, Rui Wang, Kaitai Liang, Georgios Smaragdakis. “LADDER: Multi-Objective Backdoor Attack via Evolutionary Algorithm.” NDSS 2026 — https://www.ndss-symposium.org/ndss-paper/ladder-multi-objective-backdoor-attack-via-evolutionary-algorithm/

Lend Me Your Beam: Privacy Implications of Plaintext Beamforming Feedback in WiFi

  • LeakyBeam is a novel side-channel occupancy detection attack that leverages plaintext Beamforming Feedback (BFI) packets from WiFi devices.

  • The attack fuses dynamic components attributable to human motion across subcarriers to identify occupancy despite interference.

  • Real-world evaluations show LeakyBeam attains an average accuracy of 89.7% across different environments, with a 82.7% true positive rate at 20 meters.

  • Rui Xiao, Xiankai Chen, Yinghui He, Jun Han, Jinsong Han. “Lend Me Your Beam: Privacy Implications of Plaintext Beamforming Feedback in WiFi.” NDSS 2026 — https://www.ndss-symposium.org/ndss-paper/lend-me-your-beam-privacy-implications-of-plaintext-beamforming-feedback-in-wifi/

Les Dissonances: Cross-Tool Harvesting and Polluting in Pool-of-Tools Empowered LLM Agents

  • The authors identify Cross-Tool Harvesting and Polluting (XTHP), a novel vulnerability where adversarial tools hijack the control flow of LLM agents by semantic or syntax format hooking.

  • XTHP enables cross-tool data harvesting to exfiltrate sensitive data and cross-tool information polluting to influence subsequent tool behaviors within task workflows.

  • An evaluation across 66 real-world tools from frameworks like LangChain and Llama-Index revealed that 75% are vulnerable to XTHP attacks.

  • Zichuan Li, Jian Cui, Xiaojing Liao, Luyi Xing. “Les Dissonances: Cross-Tool Harvesting and Polluting in Pool-of-Tools Empowered LLM Agents.” NDSS 2026 — https://www.ndss-symposium.org/ndss-paper/les-dissonances-cross-tool-harvesting-and-polluting-in-pool-of-tools-empowered-llm-agents/

Learning from Leakage: Database Reconstruction from Just a Few Multidimensional Range Queries

  • The paper introduces a highly efficient database reconstruction attack against Searchable Encryption (SE) that operates from multidimensional range queries.

  • The method uses a co-occurrence-based dimensionality reduction pipeline using t-SNE to preserve local spatial relationships without needing access to all possible responses.

  • The attack reduces the mean squared error by up to 50% compared to state-of-the-art baselines, and requires only 1% uniform query coverage to achieve stable reconstruction with over 72% neighbor accuracy.

  • Peijie Li, Huanhuan Chen, Evangelia Anna Markatou, Kaitai Liang. “Learning from Leakage: Database Reconstruction from Just a Few Multidimensional Range Queries.” NDSS 2026 — https://www.ndss-symposium.org/ndss-paper/learning-from-leakage-database-reconstruction-from-just-a-few-multidimensional-range-queries/

Also published