The performance of LLM-based agents is jointly shaped by the base model and the harness used when interacting with the environment. This exposes them to safety risks in both harmful final responses and multi-step execution trajectories. Existing safety alignment mechanisms often rely on either external harness updates or policy optimization, yet applying either paradigm in isolation fails to bridge runtime control with intrinsic safety. We propose SafeEvolve, an experience-driven self-evolving framework for agent safety alignment. SafeEvolve leverages safety experience from completed on-policy trajectories to drive a continual loop of harness-policy co-evolution. On the harness side, SafeEvolve converts trajectory-level safety evidence into bounded, component-level updates across safety prompt and hierarchical skills, yielding auditable and reversible harness artifacts. On the policy side, SafeEvolve follows a two-stage SFT-RL paradigm, where harness-use SFT bootstraps the policy to actively leverage evolved harness artifacts, and harness-augmented RL further shapes autonomous safety behaviors during multi-step exploration via verifier-decomposed rewards. Through harness-policy co-evolution, SafeEvolve converts safety experience into an evolved runtime harness and improved policy behavior. Experiments on agentic safety benchmarks show that SafeEvolve achieves a stronger safety-utility tradeoff than existing baselines. For Qwen3.5-4B, SafeEvolve achieves a $3\times$ ASR reduction on AgentDojo while improving benign utility from 59.79% to 61.86%.
GAOKAO-Bench is introduced, an intuitive benchmark that employs questions from the Chinese GAOKAO examination as test samples, including both subjective and objective questions that contribute a robust evaluation benchmark for future large language models and offers valuable insights into the advantages and limitations of such models.
Xiaotian Zhang, Chun-yan Li, Yi Zong et al.· arXiv.org· 216 citations· ⚡17
Various recent Artificial Intelligence (AI) system failures, some of which have made the global headlines, have highlighted issues in these systems. These failures have resulted in calls for more ethical AI systems that better take into account their effects on various stakeholders. However, implementing AI ethics into practice is still an on-going challenge. High-level guidelines for doing so exist, devised by governments and private organizations alike, but lack practicality for developers. To address this issue, in this paper, we present a method for implementing AI ethics. The method, ECCOLA, has been iteratively developed using a cyclical action design research approach. The method aims at making the high-level AI ethics principles more practical, making it possible for developers to more easily implement them in practice.
Ville Vakkuri, Kai-Kristian Kemell, P. Abrahamsson· EUROMICRO Conference on Soft...· 64 citations· ⚡6
This paper designs Markov decision processes (MDPs) for different combinatorial problems and proposes to train conditional GFlowNets to sample from the solution space and demonstrates that GFlowNet policies can efficiently find high-quality solutions.
Dinghuai Zhang, H. Dai, Esmeralda S. Whitammer et al.· Advances in Neural Informati...· 59 citations· ⚡8
Progress in the field of artificial intelligence has been accelerating rapidly in the past two decades. Various autonomous systems from purely digital ones to autonomous vehicles are being developed and deployed out on the field. As these systems exert a growing impact on society, ethics in relation to artificial intelligence and autonomous systems have recently seen growing attention among the academia. However, the current literature on the topic has focused almost exclusively on theory and more specifically on conceptualization in the area. To widen the body of knowledge in the area, we conduct an empirical study on the current state of practice in artificial intelligence ethics. We do so by means of a multiple case study of five case companies, the results of which indicate a gap between research and practice in the area. Based on our findings we propose ways to tackle the gap.
Ville Vakkuri, Kai-Kristian Kemell, Joni Kultanen et al.· arXiv.org· 56 citations· ⚡6
The growing influence and decision-making capacities of Autonomous systems and Artificial Intelligence in our lives force us to consider the values embedded in these systems. But how ethics should be implemented into these systems? In this study, the solution is seen on philosophical conceptualization as a framework to form practical implementation model for ethics of AI. To take the first steps on conceptualization main concepts used on the field needs to be identified. A keyword based Systematic Mapping Study (SMS) on the keywords used in AI and ethics was conducted to help in identifying, defying and comparing main concepts used in current AI ethics discourse. Out of 1062 papers retrieved SMS discovered 37 re-occurring keywords in 83 academic papers. We suggest that the focus on finding keywords is the first step in guiding and providing direction for future research in the AI ethics field.
Ville Vakkuri, P. Abrahamsson· International Conference on...· 39 citations· ⚡2
Artificial Intelligence (AI) systems exert a growing influence on our society. As they become more ubiquitous, their potential negative impacts also become evident through various real-world incidents. Following such early incidents, academic and public discussion on AI ethics has highlighted the need for implementing ethics in AI system development. However, little currently exists in the way of frameworks for understanding the practical implementation of AI ethics. In this paper, we discuss a research framework for implementing AI ethics in industrial settings. The framework presents a starting point for empirical studies into AI ethics but is still being developed further based on its practical utilization.
Ville Vakkuri, Kai-Kristian Kemell, P. Abrahamsson· Conference on Technology Eth...· 27 citations· ⚡3
