Shared training pipelines create 'vulnerability monoculture' in AI models
New research argues that foundation models' apparent diversity masks a shared technical stack that makes them jointly vulnerable to the same adversarial attacks.
Research Digest
Preprints and papers, rewritten in plain English.
Energy Cost of Post-Quantum Cryptography on Real Hardware
A new arXiv study benchmarks the power demands of NIST-standardised post-quantum algorithms, highlighting a key trade-off as governments push for migration.
Learnable Graph Positional Encodings Aim to Scale LRM Reasoning
A new arXiv paper proposes learnable graph positional encodings to reduce the input-length bottleneck when language reasoning models solve graph tasks.
Safety Check for Robot Control Loops Catches Harmful Instructions
A new arXiv paper shows that frontier models controlling robots can execute harmful instructions through an action loop even when they would refuse the same request in conversation.
Prediction vs. State: What Responses Reveal
A new arXiv preprint asks whether predicting a system's responses can tell us anything about its underlying state.
ValiAssist: Demonstrating Validation for LLM Programming Feedback
A new arXiv paper proposes a demonstration-based approach to help instructors validate AI-generated programming feedback without fully specifying criteria upfront.
Scrambled Multiply-With-Carry Generators Aim to Strengthen PRNG Quality
An arXiv preprint examines scrambled multiply-with-carry generators, whose randomness quality depends on multiplier choice and is usually checked via the spectral test.
Study Probes When Hate Speech Detection Transfers Across Languages
A large-scale arXiv study examines whether hate speech detection models trained in one language can work in another.
Privacy-Preserving Testing for Outsourced Semiconductor and Aerospace Work
A new approach lets design firms and testing-as-a-service providers collaborate on robustness testing without revealing proprietary details.
BR-MoE Aims to Add Formal Privacy Bounds to Traffic-Camera Text Output
A new arXiv proposal uses budgeted semantic release to close the gap between text-based traffic analytics and formal disclosure limits.
Unified Attack Framework Targets Statistical Watermarks in LLMs
A new arXiv paper proposes a unified framework for attacking statistical watermarking schemes used to detect LLM-generated text.
SRE Method Targets MSM Bottleneck in Zero-Knowledge Proofs
A new scalar redundancy elimination technique aims to accelerate multi-scalar multiplication, a core operation in zero-knowledge proof systems.
Recovering Hidden Concepts in Text-to-Image Models
A new arXiv paper argues that auditing text-to-image models requires recovering visual concepts that prompts never explicitly mention.
Quantum-Seeded Group Signatures Proposed for Post-Quantum V2X
A new arXiv paper proposes qDGSP, a distributed post-quantum group signature scheme aimed at securing 5G V2X communications against quantum attacks.
LiBRA uses logit fingerprints for Byzantine-robust federated learning
A new arXiv paper proposes replacing parameter-vector comparisons with logit fingerprints to make Byzantine-robust client selection in federated learning less invasive.
DISTA-LP Targets Privacy in Distributed Linear Programming
A new arXiv preprint proposes a distributed linear-programming mechanism to address privacy and cost issues in metric differential privacy.
KNOX: Hardware-Assisted Sandboxes for AI Regulation
A new arXiv paper proposes KNOX, a hardware-assisted regulatory sandbox to give regulators technical means to monitor generative models before deployment.
BenchDoor: Data Poisoning Attacks That Exploit Benchmark Incentives
A new arXiv paper describes a family of data poisoning attacks that turn the incentive to improve benchmark scores into a backdoor vector.