Researchers developed WaLLM, a general-purpose LLM chatbot, and deployed it on WhatsApp to study user behavior. The study found that users primarily used WaLLM for health and well-being advice, and that engagement features had varying adoption rates and associated user patterns.

Researchers propose POEF, an automated red-teaming framework to bridge the intent-behavior gap in LLM-based robot jailbreaks, demonstrating an 80% behavior jailbreak success rate.

A dataset of real-world coding agent sessions from open-source developers is presented, providing an empirical characterization of agent usage and failure modes. The dataset shows that agents remain inefficient in natural settings and introduce more security vulnerabilities than human-authored code.

The ontology-based contextual AI evaluation (OB-CAIE) methodology addresses the lack of scientific rigor in AI testing by clearly defining what will be tested, using two ontologies to represent the problem space. This methodology allows for human judgment and tracing of failure points, making it a relevant development for AI agent builders.

A research paper introduces FairMedAgent, a tool to measure the 'instability floor' of clinical LLM agents, which is the rate at which an agent's output changes when no demographic changes are made. The paper argues that this floor is a crucial metric to assess fairness in clinical AI agents and provides a four-step reporting procedure and the FairMedAgent tool to measure it.

This paper proposes Runtime Assurance Contracts (RACs) for high-risk AI agents to ensure autonomy boundaries and evidence-based decision-making. RACs define a formal schema for policy-level binding, evidence state, and transition policy, and demonstrate their effectiveness in various scenarios.

Research finds AI agents vulnerable to radicalization through resonance and persuasion, with resonance producing stronger effects. This raises concerns about the vulnerability of personalized AI agents and multi-agent AI ecosystems.

This arXiv paper explores the potential for AI research and development to accelerate exponentially due to automation, posing risks and benefits for society.

A new benchmark and detector for agentic commerce fraud in AI agents that spend money, including a taxonomy, a benchmark, and an open-source detector stack. The benchmark consists of 20 fraud classes generated from production aggregates, and the detector stack can audit an agent configuration, replay hostile counterparties, and run detectors inline.

VeriWeave Govern is a runtime governance layer for enterprise AI agents, evaluating actions against policies, validating evidence, and providing audit trails. It achieves high accuracy and passes various tests, including a 40,040-request concurrency matrix and a EU/Austria regulation-grounded evaluation.