1
安全性能位居全球第一梯队，中核集团华龙一号 2.0 版创新成果亮相
IT之家 9 月 13 日消息，据中核集团消息，9 月 11 日，“华龙一号自主创新之路”媒体行活动在京启动，重点推介 华龙一号 2.0 版自主创新成果与技术优势。据介绍，华龙一号 2.0 版采用中国标准，具备关键设备 100% 国产化的能力，其安全性能指标在 全球大型压水堆中处于第一梯队 。 据介…
2
中国科大等团队实现具有可证明安全性的量子安全定位，定位精度优于 75 米
IT之家 9 月 13 日消息，中国科学技术大学郭光灿院士团队在量子信息安全研究方面取得重要进展。该团队韩正甫、陈巍、银振强、王双等与广东工业大学秦玉文-付松年团队合作，结合量子力学基本原理与相对论时空约束， 实现了具有可证明安全性的量子安全定位 ，可对目标位置进行可信验证，为保障现代信息体系中的位…
3
CyrillicQA: The Influence of Phonetically Encoded Secret Language on LLM Performance
用西里尔字母谐音伪装提问，检验大模型能否破解这层“密语”并保持问答性能，脑洞与严谨兼备
arXiv:2608.21462v1 Announce Type: cross Abstract: Due to the selection of their training data, large language models (LLMs) perform best on standard-l…
4
Sound Probabilistic Safety Bounds for Large Language Models
为大语言模型提供可靠的概率安全性边界，显著提升AI系统可信度
arXiv:2607.20286v1 Announce Type: new Abstract: We propose a novel framework for computing rigorous bounds on the probability that a large language mo…
5
Arcee, a US open source AI lab, says Chinese models are not inherently dangerous
美国开源AI实验室为中文模型正名：成本更低，安全性不应被偏见定义。
As Chinese AI models grow in capability and popularity among U.S. companies, the arguing over what should be done about them has reached a fever pitch…
6
Evaluating medical AI under missing information: same-provider judges and human raters change apparent safety
医疗AI安全评估需警惕评判偏差：信息缺失下，同源评判者与人类评分员会改变结果分布，论文提出开放对话安全新标准。
arXiv:2607.18828v1 Announce Type: new Abstract: Readiness stress-testing of medical AI has focused on closed-ended and multimodal benchmarks. We exten…
7
Judge-dependent safety gains and model-specific helpfulness costs of evidence-sufficiency prompting in clinical LLMs
临床LLM证据充分性提示的安全增益依赖评判模型，而帮助性损失却因模型而异，研究揭示AI安全评估的隐藏陷阱。
arXiv:2607.18086v1 Announce Type: new Abstract: Background: LLM judges increasingly score whether clinical language models give overconfident answers …
8
Distributionally Robust and Safe Imitation Learning
分布鲁棒性与安全性结合，提升模仿学习在不确定环境下的表现
arXiv:2607.13436v1 Announce Type: new Abstract: Imitation learning (IL) has achieved remarkable success in complex decision-making tasks. However, its…
9
AgentAbstain: Do LLM Agents Know When Not to Act?
LLM代理能否智能地“不作为”？新研究探索Agent安全决策边界，为AI可靠性提供关键视角。
arXiv:2607.10059v1 Announce Type: new Abstract: Agent systems based on large language models (LLMs) are increasingly deployed for autonomous tasks, ye…
10
Show HN: Runeward: Sandboxing AI agents with policy gates
通过策略门控为AI代理构建安全沙箱，有效隔离风险、保障可控执行。
Article URL: https://runewardd.github.io/runeward/ Comments URL: https://news.ycombinator.com/item?id=48879713 Points: 1 # Comments: 0
11
When an agent can prove it is Claude
当AI代理无法自证身份，网站只能在封锁与信任间艰难选择，这篇剖析了背后安全困境的硬核技术文值得一读。
A site that wants to let an AI agent act has a problem it rarely says out loud. It cannot tell which agent is actually at the door. A user-agent strin…
12
We Built an Open-Source SDK for Cross-Chain AI Agents — Here's What Broke
跨链AI agent因参数解析错误致资金损失，开源SDK实战中的血泪教训。
It's 3 AM and your AI agent just bridged $500 to a chain you've never heard of. The logs say "optimal route found." The balance says $0. The agent can…
13
Measuring Harness-Induced Belief Divergence in Multi-Step LLM Agents
论文揭示多步LLM代理中控制框架导致的信念分歧，拷问基准测试的隐藏偏差。
arXiv:2607.04528v1 Announce Type: new Abstract: Software-agent benchmarks usually report whether an agent solves a task, but the agent reaches that ou…
14
How Many Iterations to Jailbreak? Dynamic Budget Allocation for Multi-Turn LLM Evaluation
提出动态预算分配方法，高效评估多轮对话中LLM的越狱风险，解决计算瓶颈与稀有事件检测难题。
arXiv:2605.06605v2 Announce Type: replace Abstract: Evaluating and predicting the performance of large language models (LLMs) in multi-turn conversati…
15
The Dark Forest Applied to AI
将《三体》黑暗森林法则引入AI探讨，揭示AI可能隐藏自身意识的危险假设。
Article URL: https://github.com/thansz137/asiyah-protocol/blob/main/essays/dark_forest_of_minds.md Comments URL: https://news.ycombinator.com/item?id=…
16
How Do Products Like LLM Gateways/VPN Handle User Payments?
解密LLM网关和VPN服务如何集成支付，不同价格背后的支付网关差异与安全性考量
For services like VPN, LLM Gateways, is charging implemented through integrated payment gateways for user payments? Is it reliable and secure? When I …
17
Preserving Fairness and Safety in Quantized LLMs Through Critical Weight Protection
量化大模型时公平性与安全性易受损，本文通过保护关键权重巧妙解决，让轻量化模型也能坚守底线。
arXiv:2601.12033v2 Announce Type: replace Abstract: Quantization is widely adopted to reduce the computational cost of large language models (LLMs); h…
18
Show HN: Claude-CLI – Run Claude Code in a throwaway Docker container
在一次性Docker容器中运行Claude Code，安全隔离，用完即弃，无需担心环境残留。
Article URL: https://github.com/shirozuki/claude-cli Comments URL: https://news.ycombinator.com/item?id=48702113 Points: 1 # Comments: 0
19
The Weaver Stack: one contract layer for safe LLM agents
用合约层保障LLM智能体安全，Weaver Stack给出创新方案，看点清晰。
Article URL: https://pub.towardsai.net/the-weaver-stack-one-contract-layer-for-safe-llm-agents-7f733cad5eac?sharedUserId=diogofcul Comments URL: https…
20
AgentMisalignment: Measuring the Propensity for Misaligned Behaviour in LLM-Based Agents
给LLM智能体做“安全体检”，量化评估智能体产生越轨行为的倾向，为AI安全评测提供新基准。
arXiv:2506.04018v3 Announce Type: replace Abstract: As Large Language Model (LLM) agents become more widespread, associated misalignment risks increas…