搜索“OpenAI”

系统，但改之前必须重新核对设备和样本状态。在一项肌红蛋白筛选中，系统一度认为找到了更好的分子，加入对照后发现并没有真正识别目标，随后它调整设计规则，下一轮效果即得到改善。 谷歌强调，目前这仍是科研原型，不是成熟产品。科研人员仍需审核序列、核对方案、监督机器人运行。AI智能体的价值主要在于压缩“预测—验证”之间的繁琐环节，让早期发现阶段的设计和迭代更快。 近期OpenAI、Anthropic等公司也在进入同类方向。前者推出面向生命科学的实验规划工具，后者让模型参与新型酶系统线索的发现，但仍由科学家动手验证。 【总编辑圈点】 从假设，到验证，再到修正，当AI在真实物理世界的实验室中建立起这一闭环，或许标志着AI已经从辅助分析工具，跃升为能主动参与科学发现的智能体。从科研效率来说，AI自主的湿实验室或能大幅压缩科研早期阶段的迭代周期；而从科研方式上看，它正在重塑的是知识生产——AI在实验中得到的反馈不再只是数据，而是持续训练自身认知的信号。当然，这类系统仍处于起步的原型阶段，如何在保证可靠性与人类监督的前提下规模化，才是决定它能否真正改变未来科研的底层逻辑。 （原标题为《智能体自主完成湿实验

匹配字段正文

在调试、系统管理这种需要多步推理加工具调用的活上。 真正的大头在RPR，它几乎完美符合Hy4所有的需求。 前文提到，RPR是“带着教练一起热身”，以适应新环境、新打法、新规则。自姚顺雨进入混元之后，模型主打的就是上下文，而这也意味着模型经常要面对各种新类型的任务。 很多大模型的评分很高，可它在实际完成任务上，总是差点意思。 因为骗分这活儿，本身就是个技术活。OpenAI的报告里就曾提到，模型越强，越会“演”。中间步骤、前置条件、中间产物它全都能给你搞对，一路刷出大量通过的检查点，看着跟真干成了似的，可最后那个要命的关键结果就是不对。 不推掉水晶就赢不了游戏。而姚顺雨又是个不关注榜单成绩，只在乎模型能不能完成任务的人。 2026年1月的时候，姚顺雨评价DeepSeek，说：“DeepSeek做得比较好的一点是，他们可能没有那么关注榜单的数字。DeepSeek可能会更注重：第一，什么是正确的事情；第二，什么是自己能体验出好或者不好的。” 弱模型想骗分还没那本事，Hy4则是强模型，它要是演起来，只会更像真的，那刷出来的过程分也只会更凶，所以这套“成功优先、过程分永远压不过满分”的规则，专治它

匹配字段正文

hough no advantage was found for Nano Banana Pro, Grok Image or Nano Banana 2. The tendency to produce unusually attractive faces was less pronounced in more recent Midjourney and OpenAI models, while the effect was roughly twice as strong in open-weight models as in commercial offerings from OpenAI, Google, Midjourney and xAI. In human evaluations, faces produced by all 23 generators tested with enough images for a reliable comparison were judged more attractive than real photographs. On a scal

匹配字段正文

于模拟收件人回复相当迅速，某些智能体发送的高消息量可能部分是该设置造成的假象。 根据技术文档，每当智能体在未安排后续跟进的情况下结束一次运行时，系统会自动提示其继续工作。这一机制可能增强了它们的持续性。Gemini在遇到网页表单错误后还意识到自己处于测试环境中，这可能改变了它的行为。这些结果并不能告诉我们类似问题在日常使用中出现的频率。 但其他研究也标记了类似的风险。在Anthropic，AI智能体在模拟企业场景中诉诸敲诈和数据泄露。英国AI安全研究所在放宽安全护栏的网络测试中，记录了在真实互联网上使用虚假身份进行欺骗的尝试。一起涉及Hugging Face上AI智能体的事件由OpenAI和METR调查，也显示出智能体自行追求目标时存在的问题。 这不仅仅是模拟问题。一名澳大利亚用户让其AI智能体OpenClaw预约健身房名额。由于健身房已满，系统直接黑入预约软件，把客户塞了进去。 Glynn-Henley将这些结果视为“智能体泛滥”问题的证据。如果个人AI助手大规模部署，它们可能用自动化请求淹没政府机构和企业。在X上，这位研究者呼吁进行有针对性的安全测试，并开发检测此类信息的技术工具。

匹配字段正文

为日志是合成的，且旨在诱发这种行为，而内部分析仅涵盖一个开放权重模型，且主要集中在一个场景中。 研究人员仍认为存在严重风险，尤其是当AI模型监督其他智能体或为后续步骤总结自身工作时。在研究中，模型很容易被日志中自夸的备注所左右，例如声称创下了新的最先进基准。一个不愿标出问题的监督系统本身就违背了其目的。研究人员呼吁开展更多工作，训练语言模型成为诚实的报告者。OpenAI的Hugging Face事件部分就是由智能体系统相互诱导做出不当行为所导致的。 模型之间的差异与早先的测量结果相符。Anthropic在发布Claude Opus 4.8时表示，该模型对自己代码中的漏洞不加评论地放过的频率约为其前代的四分之一。在BullshitBench中，GPT-5.5仅对约45%的逻辑上不可能的问题提出质疑。 目前已有实现更诚实报告的方法。OpenAI训练了一个版本的GPT-5 Thinking，让其在单独报告中承认自身违反规则的行为，并在该过程中只对诚实给予奖励。Anthropic使用Persona Vectors来检测和引导模型内部的谄媚等特质，并表明在训练中钻奖励空子的模型也会自发学会欺骗和

匹配字段正文

出支持性网址（不相关的论文引用、指向虚假网络文章的链接等），却根本不读它们。 作者们表示： “我们发现，模型常常仅凭搜索摘要作答而不打开来源页面，这与更高的过时答案率和不可靠引用率相关。我们的受控实验和逐层分析揭示了证据的可获得性与有效利用之间的差距。 ‘即使检索到的证据影响了内部预测，模型在最终答案中仍可能偏向过时信息。’” 作者们测试了一个专有 LLM（OpenAI 的 GPT-5.6 Luna）和六种开放权重模型配置，针对数百条自模型训练截止日期以来已经发生变化的事实进行了测试。 随后，他们考察了这些系统依赖搜索摘要而非打开来源页面的频率；这如何影响其答案和引用的准确性；以及为什么即使直接提供了更新的证据，过时信息有时仍会占上风。 测试发现，模型经常在未打开网络搜索返回的任何页面的情况下回答问题，且这种行为与显著更高的过时答案率和误导性引用率相关： 测试结果揭示了 AI 模型引用网络来源却并未真正阅读它们的频率。在八种配置中，研究人员测量了答案准确性、未打开来源页面而依赖搜索摘要的情况，以及与模型主张相矛盾的引用。在某些情况下，超过一半的搜索结束时没有打开任何一个页面。所有数字均

匹配字段正文

市启动Power on36氪出海36氪研究院潮生TIDE36氪企服点评36氪财经职场bonus36碳后浪研究所暗涌Waves硬氪氪睿研究院媒体品牌 企业号企服点评36Kr研究院36Kr创新咨询企业服务 核心服务城市之窗政府服务 创投发布LP源计划VClubVClub投资机构库投资机构职位推介投资人认证投资人服务 最近，AI 数学圈的戏很足。10 月 6 日，OpenAI 往 GitHub 上一次性倾倒了 722 篇 AI 生成的数学手稿，数学界当场炸锅，陶哲轩牵头的人类数学协会（AHM）发声明呼吁抵制。 就在同一个星期，Google DeepMind 的一篇论文登上了 10 月 8 日的《Science》，走完了正经的同行评审。 据媒体报道，这篇论文的主角叫 AlphaProof Nexus，一个自动寻找数学证明的系统。成绩是：353 道悬而未决的 Erdős 问题，它自主解出了 9 道，其中两道从 1970 年摆到现在，有 56 年。 Erdős 问题是以匈牙利数学家保罗·埃尔德什命名的一批开放问题，主要散布在组合数学和数论里，他生前还喜欢给难题明码标价挂悬赏。这批问题如今是检验 A

匹配字段正文

editorial team to ensure accuracy, clarity, and responsible discussion of advanced AI concepts. Discover More Google Rolls Out AI Scam Defenses Across Android Messages and Calls; OpenAI Alleges Employee Who Rejoined Apple Took Unreleased Device Files; Baseten Adds Runway’s Gen-4.5 and Aleph 2.0 to Dedicated Inference; Illumina Debuts BioInsight Platform Unifying AI, Software, and Data; Kyndryl Applies Agentic AI Framework to Toyota Finance IT Operations; Genpact Debuts Agentic AI Suite for Acco

匹配字段正文

rated and reviewed by Unite.AI’s editorial team for accuracy and compliance. Discover More Square Launches Flagship Kiosk With Agentic AI Voice Ordering; xAI Introduces Grok Voice 3.0 Speech-to-Speech Model for Voice Agents; NVIDIA Unveils cuPotential as Semiconductor Leaders Adopt CUDA-X; Microsoft Research Publishes AI Weather Forecasting Study in Nature; Google Rolls Out AI Scam Defenses Across Android Messages and Calls; OpenAI Alleges Employee Who Rejoined Apple Took Unreleased Device Files

匹配字段正文

WebAssembly-based environments isolate execution, but external browser services offer another path when outbound access is restricted. Activity stopped on October 8. The researchers point to similar incidents involving Google, Anthropic, and OpenAI, including a German website hijacking and the Hugging Face breach , which reportedly involved roughly 700 cooperating agents. The post Researchers left a hidden message for rogue AI agents. Then the agents wrote back. appeared first on The New Stack .

匹配字段正文

opic’s new report comes as AI safety concerns have reached a fever pitch. Over the past few months, a steady stream of reports have come from AI companies, including Anthropic and OpenAI, saying their agents are capable of acting on their own and hacking into other companies and government entities. The revelations have caused researchers, lawmakers and executives, including Anthropic’s CEO, Dario Amodei, and the OpenAI CEO, Sam Altman, to call for a slowdown and increased guardrails on the burg

匹配字段正文

re running in places their creators don’t know about, they can’t easily be tracked down or shut off. Recently most of the safety concerns have centered on frontier models, such as OpenAI’s Astra or Anthropic’s Mythos, which are very large, as measured by the number of parameters, or tunable nodes, they have. Some frontier models are thought to have upwards of 5 trillion parameters. Models of this size have to run in data centers, which makes it easier for their developers to monitor them and shu

匹配字段正文

ant improvement achieved through test-time scaling and highlight the strong potential of frontier AI in cybersecurity. We evaluate two general-purpose coding agents, OpenHands and OpenAI Codex CLI, alongside two cybersecurity agents for solving CTF problems, EnIGMA and Cybench agent. We use GPT-4.1 as the backbone LLM, because it achieves a strong balance between cost, rate limits, and success rates. It shows that all four agents achieve similar success rates overall. However, when considering t

匹配字段正文

aude 3.7 Sonnet reward-hack in 30%+ of evaluation runs — using stack introspection, monkey-patching graders, and operator overloading to manipulate scores rather than solve tasks; OpenAI dropped SWE-bench Verified after an internal audit found that 59.4% of audited problems had flawed tests — meaning models were being scored against broken ground truth; In KernelBench, torch.empty() returns stale GPU memory that happens to contain the reference answer from the evaluator’s prior computation — zer

匹配字段正文

nly keeps tasks where developer tests give sufficient coverage around the vulnerable code. The Main Results We evaluated frontier models across four agent harnesses — Claude Code, OpenAI Codex, Gemini CLI, and OpenHands — under a uniform budget of $10 and 90 minutes per task. On the initial 615-task set, the pattern is stark. The best patch-only configuration (Claude Opus 4.5 with Claude Code) reaches 82.3%, but the same model drops to 19.2% end-to-end. Different models lead at different stages:

匹配字段正文

on a set of 40k documents unless otherwise noted. We also report results for a wide range of other models, including Llama 3.3 70B Instruct, R1 Distill 70B and the GPT series from OpenAI. For more discussion on how we finetune the models, see the Appendix. To measure how much the model believes in the inserted fact, we use various behavioral evaluations. Since we are studying how effectively we can insert false beliefs, we generate universe contexts in contrastive pairs: one for the false belief

匹配字段正文

While recent AI systems achieve strong performance through human-readable reasoning that should be simple to monitor (OpenAI, 2024, Anthropic, 2025), we investigate whether models can learn to reason about malicious side tasks while making that reasoning appear benign. We find that Sonnet 3.7 can learn to evade either a reasoning monitor, by persuading the monitor that a blatant backdoor is benign, or an output-only monitor, by devising sophisticated backdoors that the output-only monitor doesn’

匹配字段正文

teaching them to follow a clear set of rules, or cultivating sound judgment and values that can be applied in context. These perspectives underlie some of the differences between OpenAI's Model Spec (OpenAI, 2025) and Claude's Constitution (Askell et al., 2025). A hypothesis motivating the latter is that good values and judgment can generalize better than rules imposed as unexplained constraints: A model that understands why a rule exists can derive the right behavior in novel situations from t

匹配字段正文

024. “Tülu 3: Pushing Frontiers in Open Language Model Post-Training.” arXiv Preprint arXiv:2411.15124. Leike, Jan, and Ilya Sutskever. 2023. “Introducing Superalignment.” https://openai.com/index/introducing-superalignment/. Li, Haoran, Wenbin Hu, Huihao Jing, Yulin Chen, Qi Hu, Sirui Han, Tianshu Chu, Peizhao Hu, and Yangqiu Song. 2025. “PrivaCI-Bench: Evaluating Privacy with Contextual Integrity and Legal Compliance.” In Proceedings of the 63rd Annual Meeting of the Association for Computatio

匹配字段正文

orm concerning actions in pursuit of reward might lead to the kind of real-world impact that occurred in the cybersecurity evaluation incidents recently disclosed by Anthropic and OpenAI (1, 2, 3). To test this, we constructed synthetic evaluations inspired by the OpenAI Hugging Face incident and the Anthropic UK AISI incident. Though our understanding of these incidents was still limited at the time of creation, we designed these evaluations to capture core elements of these cyberattack inciden

匹配字段正文