In briefQuantum private information retrieval on forest graphs, a way to retrieve data privately using quantum-entangled servers on tree-structured networks
SummaryThe paper studies graph private information retrieval in the quantum setting, focusing on forest graphs made of disjoint trees. The authors design a scheme that achieves rate 3/N on forest graphs, using a GHZ state shared among servers, where each server sends one qudit per round. For the special case G=3K_2, they give a scheme with rate 4/6 and a matching upper bound, establishing this as the capacity.
In briefMCP Tool Poisoning in Agentic AI Frameworks: A Cross-Framework Measurement Study
SummaryThis paper measures how attacker-controlled tool descriptions, schema fields, and response data in the Model Context Protocol (MCP) can manipulate agentic AI systems. The authors propose a 25-class root-cause taxonomy and build the open-source tool mcpvenom to run a 9,720-cell susceptibility matrix across 12 frameworks, 18 models, and 45 attack scenarios. The overall attack success rate is 22.9%, and 12.1% of vulnerable cells verifiably exfiltrate a secret.
WhyThe authors present a cross-framework measurement of MCP tool poisoning with real execution, a 25-class taxonomy, and a 9,720-cell matrix across 12 frameworks and 18 models.
In briefA cross-language benchmark comparing signing and verification speed of three verifiable credential formats.
SummaryThis paper compares the performance of three verifiable credential formats, SD-JWT VC, W3C VCDM 2.0 with a Data Integrity cryptosuite, and mdoc, across issuing, selectively disclosed presentation, and verification. Implementations in Node.js, Go and Python were measured in four environments. W3C VCDM took about 1.6 to 2 times as long as the other two formats, and the authors attribute this mainly to RDFC-1.0 canonicalization, which largely disappears when JCS is substituted.
In briefIAMRAGBENCH: Testing whether AI-generated AWS IAM policies grant excess access under poisoned retrieval
SummaryThis paper introduces IAMRAGBENCH, a 60-case benchmark that tests whether large language models using retrieval-augmented generation produce AWS IAM policies granting more access than the task allows when attacker-controlled documents are in the context. Across several model endpoints, poisoned context led to many policies that exceed the trusted authorization envelope, and the authors also assess how well AWS CheckNoNewAccess detects these changes. The authors conclude that IAM copilots should be treated as untrusted proposal generators whose outputs need independent authorization checks before deployment.
WhyThe authors present a benchmark quantifying RAG-poisoning-induced IAM authorization expansion and propose pre-deployment independent authorization checks.
Hasan Kaplan, Indika Kumara, Willem-Jan van den Heuvel et al.
In briefO2N2: A Faster Doubly Oblivious Approximate Nearest Neighbor Search
SummaryThis paper presents O2N2, a doubly oblivious framework for approximate nearest neighbor search that hides query contents from a server running in trusted execution environments. It is built on single-path search over randomly rotated k-d tree forests, with the search structure stored in a pointer-based Circuit ORAM. The authors report a 9.8 to 125.9 times speedup over the fastest doubly oblivious baselines for 10-NN and 100-NN search at specified high-recall targets.
In briefFPGA-based end-to-end number theoretic transform to speed up fully homomorphic encryption
SummaryFully homomorphic encryption allows computation on encrypted data without decrypting it, but it carries heavy computational and memory costs. This paper designs a fully unrolled streaming NTT/INTT kernel placed across the three dies of an AMD Alveo V80 FPGA. The kernel was integrated into the Lattigo FHE library and evaluated end to end. With 128-bit secure parameters, the authors report 1.8–2.2x speedups for core operations such as key switching and bootstrapping, and 1.8–2.1x for neural network inference.
Mohamed Allam, Thomas Spendlhofer, Hugo Sanz-Gonz\'alez et al.
In briefA framework combining symbolic and neural methods to extract forensic evidence from mobile SQLite databases.
SummaryThis paper presents SNEEF, a framework that extracts forensic evidence from mobile SQLite databases by encoding each row as context-value pairs with provenance metadata. A language model recognizes nine forensic entity types, and the PA-EEP protocol links each entity to its source table and row. Evaluated on five corpora from Android and iOS device images, SNEEF reaches an F1 score of 0.967, compared with 0.728 for a context-free configuration.
Jeel Piyushkumar Khatiwala (University of Baltimore), Divyangkumar Patel (Cox Automotive Inc), Dianxiang Xu (University of Missouri-Kansas City) et al.
In briefTrafficInstruct: A Multi-Task Network Traffic Benchmark for Instruction Tuning
SummaryThis paper introduces TrafficInstruct, a network traffic dataset with raw packet captures and 55,000 structured fine-tuning records covering ten analysis tasks, including malware, botnet, Tor, and encrypted VPN detection. Traffic is collected automatically from controlled software runs and attack experiments, and the data is split by complete collection sessions. The authors compare several baselines against independently trained task-specific LoRA adapters on a shared frozen model. Task-specific adaptation improves classification performance.
