主题路线Topic route
安全与合规,按顺序找资料Safety and compliance: resources in order
安全不是加一句「请勿有害」:先定威胁模型,再看评测与事件记录,然后落到权限、数据与日志。合规也不是一张清单,而是把适用法规、证据与责任人分开写清。Safety is not one added sentence: define the threat model, look at evaluations and incident records, then implement permissions, data handling and logs. Compliance is not one checklist either; separate the applicable rules, the evidence and the owner.
路线顺序不代表推荐排名;没有实测的能力对比、成本降幅与成功率,站群一律写未知。Route order is not a ranking; where no measured comparison, cost reduction or success rate exists, the sites record it as unknown.
先做什么Where to start
按环节找全站资料Find material by stage
- NIST AI RMF、OWASP LLM Top 10、MITRE ATLAS、AI Incident DatabaseNIST AI RMF, OWASP LLM Top 10, MITRE ATLAS, AI Incident Database
- 评测卡的污染、复现与「分数不能横比」说明Benchmark cards on contamination, reproduction and non-comparability
- 可解释性与智能体失败分析:探针、参数分解与边界Interpretability and agent-failure analysis: probes, parameter decomposition, limits
边界Boundary
本站不是法律意见,也不替任何组织做合规结论;法规原文、监管解释与合同条款优先。安全评测分数只在对应威胁模型与测试集下成立,不能拿一个基准证明「整体安全」。This is not legal advice and does not issue a compliance conclusion for any organisation; statutes, regulator interpretations and contracts govern. Safety scores hold only for their threat model and test set; one benchmark cannot prove “safe overall”.