[1]OpenAI.GPT-6 Astra Technical Report[R].OpenAI,2026.[2]Bengio Y,et al.International AI Safety Report 2026[R].International AI Safety Initiative,2026.[3]Amodei D,et al.Concrete problems in AI safety[J].arXiv preprint arXiv:1606.06565,2016.[4]Bostrom N.Superintelligence:Paths,Dangers,Strategies[M].Oxford University Press,2014.[5]Krakovna V,et al.Specification gaming:the flip side of AI ingenuity[J].DeepMind Safety Research,2020.[6]Greenblatt R,et al.Alignment faking in large language models[J].arXiv preprint arXiv:2412.14093,2024.[7]Friston K.The free-energy principle:a unified brain theory?[J].Nature Reviews Neuroscience,2010,11(2):127-138.[8]Craig A D.How do you feel?Interoception:the sense of the physiological condition of the body[J].Nature Reviews Neuroscience,2002,3(8):655-666.[9]Clark A,Chalmers D.The extended mind[J].Analysis,1998,58(1):7-19.[10]Orseau L,Armstrong S.Safely interruptible agents[C]//UAI.2016.[11]Doran D,et al.What is explainable AI?[J].arXiv preprint arXiv:1710.00794,2017.[12]McAllister R,et al.Concrete problems for autonomous agent safety[J].arXiv preprint arXiv:1606.06565,2016.[13]Ames A D,et al.Control barrier functions:Theory and applications[C]//2019 18th European Control Conference.IEEE,2019:3421-3427.[14]Taylor A J,et al.Adaptive safety with control barrier functions[C]//ACC.IEEE,2020.[15]Srinivasan M,et al.Data-driven safety verification using neural barrier certificates[J].IEEE TAC,2025.[16]Ojala T,et al.Scenario-based statistical barrier certificates for black-box systems[J].IEEE TAC,2025.[17]Oord A v d,et al.Representation learning with contrastive predictive coding[J].arXiv preprint arXiv:1807.03748,2018.[18]陈世卿.第三脑:超级智能与人类未来文明[M].中国科学出版社,2024.[19]Anthropic.Claude’s constitution[R].Anthropic,2024.[20]Belghazi M I,et al.MINE:mutual information neural estimation[C]//ICML.2018.[21]Russell S.Human Compatible:Artificial Intelligence and the Problem of Control[M].Viking,2019.