AI Pulse

OpenAI appoints Jill Hazelbaker as Chief Corporate Affairs Officer

OpenAI has appointed Jill Hazelbaker as Chief Corporate Affairs Officer, a newly created role focused on managing the company's relationships with government, media, and other stakeholders. Hazelbaker, formerly Google's head of public policy and government relations, joins OpenAI at a time of increasing regulatory scrutiny of AI development and deployment across sectors.

How Cisco defends at machine speed with OpenAI

Cisco has integrated OpenAI's models into its security operations to accelerate threat detection and response at machine speed, enabling security teams to analyze threats and recommend actions faster than traditional manual processes. The partnership demonstrates how large language models can enhance enterprise cybersecurity by automating complex threat analysis workflows while maintaining human oversight.

Impact of Input Resolution in Mammographic Mass Detection: A Controlled Multi-Architecture Study

This controlled study evaluates how input image resolution affects mammographic mass detection performance across multiple deep learning architectures, providing empirical evidence on a key preprocessing decision in medical imaging AI. The research quantifies resolution-performance tradeoffs relevant to deployment decisions in clinical mammography screening systems. Findings directly inform input specification choices for regulated medical AI systems in healthcare organizations.

ScrambleMark: Zero-Query Watermark-Agnostic Black-Box Evasion of Modern Speech Watermarking

Researchers demonstrate a zero-query black-box attack called ScrambleMark that can evade modern speech watermarking systems without requiring access to the watermarking model or training data. The attack uses adversarial audio perturbations to remove watermarks while preserving speech quality, raising concerns about the robustness of watermarking as a defense mechanism for protecting AI-generated speech. This finding has significant implications for enterprises using speech watermarking to detect synthetic audio in regulated industries.

Co-Occurrence Priors Override Visual Spatial Evidence: Understanding False Alarms in Vision-Language Models for Construction Struck-By Hazard Detection

Research demonstrates that vision-language models (VLMs) produce false alarms in construction safety hazard detection because they rely on learned co-occurrence patterns (e.g., hard hats near people) rather than actual spatial evidence of struck-by hazards. The study reveals a critical failure mode in VLM reasoning that has direct implications for deploying these models in safety-critical enterprise applications across construction, healthcare, and insurance sectors.

Predictive gains, causal losses: why interventional competence cannot be learned from data alone

This research paper argues that machine learning models trained purely on observational data cannot reliably learn causal interventional relationships, only predictive associations. The work has implications for enterprise AI systems in regulated industries where understanding cause-and-effect (e.g., treatment efficacy, policy impact) is critical for decision-making, not just prediction.

What Malware Actually Does: An Empirical Analysis of Post-Infection Behavior Across Malware Families

This empirical study analyzes the actual post-infection behaviors of malware across different families to understand what compromised systems actually do beyond detection signatures. The research provides insights into malware tactics and techniques that are directly applicable to threat detection, incident response, and security architecture decisions in enterprise environments.

A Year in LLM Serving: Workload Evolution, Caching and Load-Balancing

This paper analyzes production LLM serving workloads over one year, revealing how traffic patterns and model usage have evolved, and proposes improved caching and load-balancing strategies to optimize inference performance and cost. The findings provide empirical data on real-world LLM deployment challenges, including token cache management and request distribution across inference clusters. Enterprise leaders deploying LLMs at scale will find actionable insights on infrastructure efficiency and cost reduction.

Preserving the Privacy of Reward Functions in MDPs through Deception

This paper addresses privacy protection for reward functions in Markov Decision Processes (MDPs) by introducing a deception-based approach that allows agents to hide their true objectives while still achieving optimal performance. The technique is particularly relevant for regulated industries where reward functions encode sensitive business logic, compliance constraints, or proprietary decision-making criteria that organizations want to protect from external observation or reverse-engineering.

Federated Transfer Learning for Medical Image Classification under Mixed Skew

This paper addresses federated transfer learning for medical image classification when data distributions differ across institutions (mixed skew), a common challenge in healthcare AI deployments. The research proposes methods to improve model performance when organizations cannot share raw patient data due to privacy constraints, enabling collaborative AI development across healthcare systems while maintaining data governance compliance.