Joint Pixel-Prompt Optimization Targets VLM Resource Exhaustion
A new arXiv paper proposes jointly optimizing image pixels and text prompts to broaden resource-exhaustion attacks on vision-language models beyond unimodal threat models.
Research Digest
Preprints and papers, rewritten in plain English.
Grounding Confusable Movie Moments in Full-Length Films
A new arXiv preprint tackles the problem of localizing a specific movie moment when many scenes share the same characters, locations, and actions.
Cross-Modal Jailbreak Attack Splits Harmful Intent Across Text and Image
Researchers identify a new safety risk in image editing models where harmful instructions are distributed across text and visual inputs to evade detection.
ContraSafe-VL Dataset Targets Multimodal Safety Alignment
A new contrastive safety preference dataset aims to help vision-language models handle safety risks that arise from the interplay of text and images.
Adversarial Training Aims to Improve AI-Generated Video Detection
A new arXiv paper proposes adversarial training to help detectors generalize beyond the generators they were trained on.
Privacy-Preserving ML for Internet of Vehicles: A Survey
A new review maps the state of privacy-preserving machine learning in connected vehicles, highlighting the tension between intelligent services and data protection.
STRAL: How Adaptive Screening Shifts SOC Alert Trade-Offs
A new active-learning method, STRAL, quantifies how adaptive alert screening changes the balance between false-alert burden and missed detections.
Devilray Baseline Reveals Blind Spots in Fake Base Station Detection
A new reconfigurable adversarial baseline aims to expose weaknesses in fake base station detection, which has been hampered by limited access to commercial devices.
Bit-Flip Attack Can Hijack LLM Decisions
New research shows that flipping bits in open-source LLMs can alter their decision-making in high-stakes contexts.
COPEX Benchmark Tests LLM Defenses Across MCP Tool-Use Layers
A new arXiv benchmark probes how well large language models resist adversarial input at each layer of Model Context Protocol tool use.
Beyond the Balance Sheet: Social Cost of Data Breaches
A new arXiv paper argues that measuring data breaches only by corporate losses undercounts the true harm to society.
Interface Attacks Can Tilt System One Model Judgments
A new paper introduces DICE, a method that exploits the interface between System One models and their applications to influence decisions.
Empirical study asks whether ERC-8004 can make agent-to-agent trust work
A new arXiv paper empirically examines the ERC-8004 protocol, which aims to help autonomous agents judge the trustworthiness of unknown counterparts.
LLM Agent Labels in Shared Context Are Not a Trust Boundary
A new arXiv paper argues that labels inside a shared context cannot be relied on as an enforceable security boundary for LLM agents.
Quantum Trick Cuts Communication in Secure Computations
A new theoretical study shows how quantum cryptography can reduce the communication cost of classical secure computations with long outputs.
Topology-Based Client Descriptors Guide Personalised Federated Learning
A new approach uses persistent-homology descriptors to decide which clients should share models in personalised federated learning.
Genomic Language Models Can Memorize DNA, Raising Privacy Risks
A new study quantifies how genomic language models may memorize training sequences, highlighting privacy concerns for sensitive genetic data.
MemPoison attacks show LLM memory can be poisoned through normal chat
A new arXiv study demonstrates that adversarial content can hide in an agent's persistent memory and corrupt later decisions.