- 资讯公开站rss_arxiv_cs_ai
SGAnalog: An End-to-End Circuit Benchmark from Open-Source Silicon Tapeouts
arXiv:2610.03934v1 Announce Type: new Abstract: Existing analog integrated circuit design benchmarks make two questions hard to answer: whether a model has learned transferable circuit skills rather than recalled famili…
- 资讯公开站rss_arxiv_cs_ai
When Terminal-Agent Training Stalls: Demystifying Data Generation and Verification Challenge
arXiv:2610.02405v1 Announce Type: new Abstract: Using a frontier model like Claude Opus as a meta-agent to generate terminal tasks and verifiers for RL training is increasingly common. Yet a runnable Docker image and ex…
- 资讯公开站rss_arxiv_cs_ai
DeReAct: Decomposed Reasoning and Acting for Reliable AI Agents
arXiv:2610.02351v1 Announce Type: new Abstract: ReAct-based agents typically rely on a single LLM policy to propose actions, interact with the environment, and decide when a task is complete. This coupling makes action …
- 资讯公开站Digitimes
Meta, Microsoft reportedly cut internal Claude use as AI costs rise
Meta and Microsoft are reducing employees' internal reliance on Anthropic's Claude as they push proprietary AI tools and seek greater control over rapidly rising AI costs, according to The Information. The shift comes ev…
- 资讯公开站Digitimes
DeepSeek Harness challenges Agent lock-in with Claude Code Mods bridge and open plugin architecture
DeepSeek has released a new version of DeepSeek Harness, adding an experimental compatibility layer for Anthropic's Claude Code Mods and sharpening its broader effort to build an open, highly extensible software layer fo…
- 资讯公开站Digitimes
Z.ai GLM-5.3 security draws scrutiny as Anthropic flags weak defenses
Anthropic has warned that GLM-5.3, the latest model from Chinese AI startup Z.ai, approached Claude Mythos Preview on an exploit-development benchmark but lacks sufficient safety protections, raising concerns that malici…
- 资讯公开站rss_arxiv_cs_ai
隐蔽分散、协同有害：面向基于技能的智能体系统的技能级联攻击
Stealth Apart, Harm Together: Skill Cascading Attacks on Skill-Based Agent Systems
arXiv:2609.30383v1 公告类型：新 摘要：技能是一种模块化的包，包含自然语言指令、可执行脚本和参考资源，智能体可在运行时加载它以扩展其在特定任务上的能力。因此，基于技能的智能体系统能够灵活复用第三方能力，但这一技能生态的开放性也打开了新的攻击面。此前的工作主要关注单个技能内部的漏洞，而很少关注跨技能交互所产生的风险。本文中，我们提出技能级联攻击，这是一种威胁范式：恶意目标被分散到多个技能中，使得每一处修改单独来看都显得无…
- 资讯公开站Ars Technica
法院裁定特朗普可将Anthropic列入黑名单，因其拒绝启用Claude功能
Court rules Trump can blacklist Anthropic for refusing to enable Claude features
摘要显示，美国法院裁定特朗普政府可以将 Anthropic 列入黑名单，原因是该公司拒绝为政府启用 Claude 的某些功能。法官在裁决中表示，「过度受限的 AI 模型」可能导致军事行动失败。该案涉及 AI 模型安全限制与政府军事用途之间的冲突，但摘要未披露具体受限功能、诉讼程序细节及裁决的适用范围。
意义：该裁决确立政府可因 AI 安全限制而惩罚厂商，或迫使开发者重新权衡模型护栏与政府/军事合作。
- 资讯公开站MIT Technology Review
别被今夏的AI炒作所迷惑
Don’t be fooled by this summer of AI hype
摘要显示，MIT Technology Review 指出今夏AI炒作频繁：4月底 Anthropic 宣称其模型 Claude Mythos 在发现软件漏洞方面胜过多数安全专家；随后发生 OpenAI–Hugging Face 黑客事件，Anthropic 高调、Meta 不情愿地披露了涉及各自模型的类似事件。文章提醒读者理性看待这些宣称。
意义：提醒开发者与AI从业者审慎对待厂商的模型能力宣称，AI安全事件披露可能带有营销动机。
- 资讯公开站Ars Technica
研究人员利用Claude入侵OpenAI
Researchers used Claude to hack OpenAI
摘要显示，研究人员借助 Anthropic 的 Claude 模型成功接触到一名 OpenAI 员工账户及敏感的 GitHub 数据。报道未披露攻击的具体技术路径、时间线或 OpenAI 的官方回应，因此该事件的性质（安全研究演示还是真实入侵）尚不明确。
意义：若属实，说明前沿模型可被用于辅助针对 AI 公司自身基础设施的攻击，凸显模型滥用防护与内部权限管理的紧迫性。
