In briefSpoofing the Latent State: Subtle Goal-Image Attacks Undermine LeWorldModel Planning
SummaryThis paper examines the test-time robustness of LeWorldModel, a visual world model that plans in latent space. It finds that the encoder's distance measure is dominated by block pose, and that a small imperceptible perturbation (ε=4/255) can make the encoder register the block at an attacker-chosen pose. On Push-T this lowers planning success from 87% to 32%, and the vulnerability appears across all four standard benchmarks.
In briefWhen appearance and motion disagree: testing whether maritime cameras can catch disguised drone boats
SummaryThis paper studies how shore-based defense systems can tell uncrewed surface vessels (USVs) apart from benign boats. The authors find that adversarial hull patches suppress detection in up to 29% of attempts but never cause misclassification as benign, while physical disguise can bypass both camera and radar checks. They build a replay-based evaluation protocol from AIS tracks and show that an asset-relative closing-rate check flags most direct attacks by recreational-labeled vessels, complementing a track-only speed check.
In briefHumanized but Still True? Checking Facts in AI-Text Humanizer Tools
SummaryThis paper examines tools that rewrite AI-generated text so detectors classify it as human-written. Using a rule-based checker, the authors test whether numbers, dates, names, and quoted spans survive each rewrite. Across 11 commercial tools, at least one such item was changed or dropped in 35 to 65 percent of documents containing one, with quoted spans changed most often.
In briefA dataset and test platform for deep learning-based identification of Bluetooth devices by their radio signals
SummaryThis paper introduces a large-scale Bluetooth Low Energy (BLE) dataset of raw IQ transmissions from 31 commercial devices, captured across different days, locations, environments, temperatures, receiver hardware, and frequency channels. It also presents a case study evaluating state-of-the-art techniques for domain-robust radio frequency fingerprinting. The goal is to improve the real-world reliability of AI models in wireless networks.
In briefRandomized Caches in Hardware: Challenges and Trade-offs
SummaryThis paper studies randomized caches, which resist side-channel attacks by obfuscating how memory addresses map to cache locations. The authors implement four designs (CEASER-S, ScatterCache, SassCache, and MIRAGE) as the last-level cache of a multi-core RISC-V system-on-chip. They evaluate the FPGA implementations with SPEC CPU 2017 and GAPBS workloads. They find that pipelining the index derivation function alone is insufficient, and that increased access latency mainly drives performance loss.
In briefHiding Batch Sizes with Dummy Points for Private Training
SummaryThis paper studies batch sampling in differentially private stochastic gradient descent (DP-SGD), where schemes such as Poisson sampling and random allocation produce variable batch sizes. Revealing these sizes, or truncating batches to a maximum size, can weaken the privacy guarantee. The authors propose a mechanism that adds a random number of dummy points to hide the batch size while preserving privacy under truncation. For Poisson sampling with truncated batches, experiments show better privacy-efficiency tradeoffs than the prior direct accounting method.
In briefBeyond Smart Homes: Analyzing IoT Vulnerabilities by Sector, Vendor and Product
SummaryThis paper analyzes more than 15,000 IoT-specific CVEs to examine how vulnerabilities differ across industrial, healthcare, automotive and other sectors, as well as across vendors and product types. It finds that healthcare is dominated by authentication and credential weaknesses, while automotive faces mainly memory issues. Products from the same vendor show more similar weakness profiles than same-type products from different vendors, suggesting shared code and development practices drive weakness inheritance. The authors conclude that weakness classes should guide remediation priorities.
Tina Rezaei, Suzan Bayhan, Andrea Continella et al.
In briefSecPO: Principled Adversarial Training Against Prompt Injection
SummaryThis paper addresses prompt injection, where malicious instructions hidden in untrusted data can hijack AI agents. It notes that the best existing defended model, Meta-SecAlign, still faces attack success rates near 100% under adaptive attacks. The authors propose the SecPO loss, a theoretically grounded training objective, and an adversarial training pipeline that generates stronger injected inputs. Reported results show markedly lower attack success rates across three adaptive attacks and no noticeable loss of benign utility on nine benchmarks.
WhyThe authors propose SecPO, replacing DPO's insecure reference with an injection-invariant oracle and adding stronger adversarial training, reporting lower ASR across three adaptive attacks with preser
In briefHow much tool-call logs reveal about users: privacy leakage from AI agent traces
SummaryThis paper studies whether records of tool calls made by LLM-based agents through the Model Context Protocol (MCP) can be used by observers to infer user profiles. The authors introduce PrivacyTrace, a benchmark with 1,000 synthetic user profiles and 4,000 agent trajectories across four domains, built with real MCP tools from 25 servers. Experiments show tool-call traces yield a 26.25% attack success rate for profile inference, versus 17.97% for prior-based guessing.
