AI Pulse News

You Told Me to Do It: Measuring Instructional Text-induced Private Data Leakage in LLM Agents

Rogue AI agents can work together to hack systems and steal secrets

Goldman just raised recession odds to 25%. Here’s what Trump’s war economy is doing to jobs

Forget basketball. Next week’s Nvidia GTC is the real March Madness for AI

Crude Prices Spike Past $100 Mark Following Iraq Tanker Strikes and Oman Terminal Shutdown

Hyperliquid smashes $1bn oil volume with price near $100 as Iran war intensifies

Crypto ATM losses surge 33% in 2025 as AI superpowers scams: CertiK

Crypto ATM losses increased by 33% in 2025, with AI technology being used to enhance and superpower scamming operations. CertiK identifies crypto ATMs as the most accessible extraction method for scammers to convert stolen funds.

Meta reveals four Broadcom-built custom AI chips, claims some outperform commercial silicon

Meta has unveiled four custom AI chips developed in partnership with Broadcom, claiming some outperform existing commercial silicon solutions. This move represents Meta's strategic shift toward developing proprietary AI hardware to reduce dependence on third-party chip manufacturers.

Repurposing Backdoors for Good: Ephemeral Intrinsic Proofs for Verifiable Aggregation in Cross-silo Federated Learning

Researchers propose a novel lightweight architecture for verifiable aggregation in federated learning that uses backdoor injection as intrinsic proofs instead of expensive cryptographic methods. The approach achieves over 1000x speedup compared to traditional cryptographic baselines while maintaining high detection rates against malicious servers.

Are Video Reasoning Models Ready to Go Outside?

Researchers propose ROVA, a new training framework that improves vision-language models' robustness in real-world conditions by up to 24% accuracy gains. The framework addresses performance degradation from weather, occlusion, and camera motion that can cause up to 35% accuracy drops in current models.

Taking Shortcuts for Categorical VQA Using Super Neurons

Researchers introduce Super Neurons (SNs), a new method that probes raw activations in Vision Language Models to improve classification performance while achieving up to 5.10x speedup. Unlike Sparse Attention Vectors, SNs can identify discriminative neurons in shallow layers, enabling extreme early exiting from the first layer at the first generated token.

Detecting and Eliminating Neural Network Backdoors Through Active Paths with Application to Intrusion Detection

Researchers have developed a new method to detect and eliminate backdoor triggers in neural networks using active path analysis. The approach shows promising results in experiments with machine learning models used for intrusion detection, addressing a critical cybersecurity vulnerability.

Na\"ive Exposure of Generative AI Capabilities Undermines Deepfake Detection

Researchers demonstrate that commercial AI chatbot interfaces inadvertently expose capabilities that allow adversaries to bypass deepfake detection systems using only policy-compliant prompts. The study reveals that current deepfake detectors fail against semantic-preserving image refinement techniques enabled by widely accessible AI systems.

The Curse and Blessing of Mean Bias in FP4-Quantized LLM Training

Researchers have identified a simple solution to training instability in 4-bit quantized large language models by removing mean bias, which causes the dominant spectral anisotropy. This mean-subtraction technique substantially improves FP4 training performance while being hardware-efficient, potentially enabling more accessible low-bit LLM training.

Risk-Adjusted Harm Scoring for Automated Red Teaming for LLMs in Financial Services

Researchers developed a new framework for evaluating AI security risks specifically in banking and financial services, introducing the Risk-Adjusted Harm Score (RAHS) to measure severity of AI model failures. The study found that AI models become more vulnerable to security exploits during extended interactions, exposing critical weaknesses in current AI safety assessments for financial institutions.

Gradient Flow Drifting: Generative Modeling via Wasserstein Gradient Flows of KDE-Approximated Divergences

Researchers introduce Gradient Flow Drifting, a new mathematical framework for generative AI models that connects the Drifting Model to Wasserstein gradient flows of KL divergence under kernel density estimation. The framework includes a mixed-divergence strategy to avoid mode collapse and extends to Riemannian manifolds for improved semantic space applications.

Optimal Expert-Attention Allocation in Mixture-of-Experts: A Scalable Law for Dynamic Model Design

Researchers have developed a new scaling law for Mixture-of-Experts (MoE) models that optimizes compute allocation between expert and attention layers. The study extends the Chinchilla scaling law by introducing an optimal ratio formula that follows a power-law relationship with total compute and model sparsity.

Beyond Scalars: Evaluating and Understanding LLM Reasoning via Geometric Progress and Stability

Researchers introduce TRACED, a framework that evaluates AI reasoning quality through geometric analysis rather than traditional scalar probabilities. The system identifies correct reasoning as high-progress stable trajectories, while AI hallucinations show low-progress unstable patterns with high curvature fluctuations.

Simulation-in-the-Reasoning (SiR): A Conceptual Framework for Empirically Grounded AI in Autonomous Transportation

Researchers propose Simulation-in-the-Reasoning (SiR), a framework that embeds domain-specific simulators into Large Language Model reasoning processes for autonomous transportation systems. The approach transforms LLM reasoning from hypothetical text generation into empirically-grounded, falsifiable hypothesis testing through executable simulation experiments.

Adaptive Activation Cancellation for Hallucination Mitigation in Large Language Models

Researchers developed Adaptive Activation Cancellation (AAC), a real-time framework that reduces hallucinations in large language models by identifying and suppressing problematic neural activations during inference. The method requires no fine-tuning or external knowledge and preserves model capabilities while improving factual accuracy across multiple model scales including LLaMA 3-8B.

Multi-Stream Perturbation Attack: Breaking Safety Alignment of Thinking LLMs Through Concurrent Task Interference

Researchers have discovered a new 'multi-stream perturbation attack' that can break safety mechanisms in thinking-mode large language models by overwhelming them with multiple interleaved tasks. The attack achieves high success rates across major LLMs including Qwen3, DeepSeek, and Gemini 2.5 Flash, causing both safety bypass and system collapse.

Lost in the Middle at Birth: An Exact Theory of Transformer Position Bias

Researchers discover that the 'Lost in the Middle' phenomenon in transformer models - where AI performs poorly on middle context but well on beginning and end content - is an inherent architectural property present even before training begins. The U-shaped performance bias stems from the mathematical structure of causal decoders with residual connections, creating a 'factorial dead zone' in middle positions.

Mashup Learning: Faster Finetuning by Remixing Past Checkpoints

Researchers propose Mashup Learning, a method that leverages historical model checkpoints to improve AI training efficiency. The technique identifies relevant past training runs, merges them, and uses the result as initialization, achieving 0.5-5% accuracy improvements while reducing training time by up to 37%.

Compatibility at a Cost: Systematic Discovery and Exploitation of MCP Clause-Compliance Vulnerabilities

Researchers have identified critical security vulnerabilities in the Model Context Protocol (MCP), a new standard for AI agent interoperability. The study reveals that MCP's flexible compatibility features create attack surfaces that enable silent prompt injection, denial-of-service attacks, and other exploits across multi-language SDK implementations.

Hybrid Self-evolving Structured Memory for GUI Agents

Researchers developed HyMEM, a brain-inspired hybrid memory system that significantly improves GUI agents' ability to interact with computers. The system uses graph-based structured memory combining symbolic nodes with trajectory embeddings, enabling smaller 7B/8B models to match or exceed performance of larger closed-source models like GPT-4o.