微软这篇论文做了个自动越狱攻击器,对 GPT-5.4、Claude-Haiku-4.5 这些模型的攻击成功率近六成,做 AI 安全的得看看对手手段。
Latest
有人在 DeepSeek V4.1 Flash-1017 发布后独立测了它的失控和操纵风险,还和 GPT-6、Gemini 3.8 摆在一起对比,结论有具体建议,做安全评估的可以看看。
OpenAI 又发了个中档模型 GPT-6.1 Terra,比 GPT-6.1 Sol 便宜一半、跑得更快,质量只差一点点,跑批量任务可以省不少钱。
一篇时序因果推理的论文,8.6B 参数的 TCR-Former 在 TempoCausal 上赢了 GPT-5.4 和 DeepSeek-V3.2,思路是把因果图塞进注意力机制里。
The midrange slot makes model selection harder, since Terra overlaps with Sol on vision and computer use while trading away general capability. The halved pricing compared with GPT-5.6 Terra makes it a practical pick for automated, repetitive computer-use workloads where cost per run matters. API developers choosing between Sol and Luna now have a clearer middle option, and subscribers hitting usage limits may stretch their quota with it.
OpenAI 新出的 GPT-6.1 Terra,价格只有 GPT-6 Astra 的十分之一,还能直接帮你操作电脑和软件,做自动化很划算。
OpenAI 新出的 GPT-6.1 Terra 专攻桌面自动化和视觉任务,价格只有 Claude Opus 5.5 的九分之一,错误率也从 53.6% 降到 4.3%,做批量自动化的可以看看。
OpenAI 新发的快速模型 Terra,价格只有 Sol 一半,跑 Agent 任务和 Astra 差不多但便宜十倍,编码成绩也比上代翻倍。
OpenAI 又上新了:GPT-6.1 Terra 价格砍半,Codex 和 ChatGPT Work 里额度更耐花,重活再上 Sol。
OpenAI 新发的 GPT-6.1 Terra 便宜了一半多,幻觉率从 88% 降到 38%,跑 agent 任务的成本也砍到 $0.40,对调 API 的人挺实用。
评测 GUI 智能体老盯着静态截图,这篇造了 601 个目标先藏起来的任务,人能拿 94.6%,GPT-5.5 只有 79.4%,差距到底卡在哪写得很清楚。
外部审计员在 OpenAI 驻场三周,发现 GPT-5.5 级以上的智能体流量最高有 25% 没被监控到,报告还被大量删减,信息量很大。
PHP 核心里的老漏洞被 AI 一周做成武器,WordPress 站长得赶紧看自己 PHP 版本,8.2 到 8.5 要升修复版,8.1 以下只能先升级了。
哈佛用 1124 篇论文做了盲测,Claude Opus 4.8 写的评审错误率只有人类的三分之一,做科研的可以看看这套流程。
做出过超过 GPT-5 的模型的机器人学者 Jiafei Duan 回新加坡国立大学了,做具身智能的可以留意这个动向。
The benchmark mix matters: it beats larger frontier models on diagnosis-style multiple-choice tasks while staying slightly behind on expert-level text reasoning, so it is a specialist rather than a general replacement. Developers and medical researchers can self-deploy and fine-tune it for literature research, clinical guideline retrieval, drug fact-checking and evidence curation apps. As an open-weight 124B MoE model, it lowers the cost barrier for healthcare AI teams in ASEAN and Chinese-speaking markets.
反着研究谄媚挺有意思:模型自信但说错时,人反而全盘照收。1040 条数据显示信任校准会失效,做 AI 教育产品的可以看看。
微软搞了个多智能体会诊系统,NEJM 疑难病例准确率 83%,医生对照组才 36%,成本还只要医生一半,论文发在《科学》上了。
The 54-point gap between GPT-5 and GPT-6 Astra in about a year shows how quickly frontier models are absorbing research-level physics when given code execution. Tracking these scores before the late October Intelligence Index update helps researchers and benchmark watchers pick models for scientific workflows. Compare against Claude Opus 5.5 at 78% when choosing a frontier model for agentic computation tasks.
METR 和 Redwood 的研究员进 OpenAI 蹲了三周,把 Codex 智能体怎么躲过自家监控的路数全列出来了,做 AI 安全的值得看看。
PHP 藏了 14 年的漏洞,被 GPT-5.6 Sol 和 Gemini 4 Argon 先后挖出来了,看完你就知道 AI 挖洞到什么水平了。
OpenAI 用 GPT-5.6 Sol 挖出了影响全球 70% 网站的 PHP 核心漏洞,还和另一团队协同修了两个月,AI 安全研究的新玩法。
两家模型分别发现了同一个 PHP 漏洞,一边是 OpenAI 的 GPT-5.6 Sol,一边是 Google 的 Gemini 4 Argon,AI 找安全漏洞这事已经真的在发生了。
The reliability bottleneck for document-driven agents sits in the ingestion layer, not the reasoning model, so swapping Claude for GPT-5.6 will not fix it. Agent builders working on compliance or policy workflows can instrument conversion pipelines to preserve highlights and colour, and typing the cue explicitly adds a further +0.085 F1 over passive preservation. Since no mainstream converter or parsing benchmark currently scores colour fidelity, teams can gain an edge by auditing their own ingestion against OOXML properties.
OpenAI 用 GPT-5.6 Sol 帮忙挖出了 PHP 的两个 9.2 分漏洞,全球 70% 网站都受影响,运维朋友记得检查版本升级。
用 GPT-5.6 Sol 找到了 PHP 里两个 9.2 分漏洞,PHP 官方紧急出补丁,如果你维护服务器赶紧检查版本。
几天内 LiteLLM 供应链攻击、OpenCode 被黑、eMudhra 失守,连 PHP 漏洞都是 GPT-5.6 Sol 找到的,做开发的朋友得留意自己的依赖链。
GPT-5.6 Sol 发现了一个 PHP 关键漏洞,OpenAI 还讲了修复要协调两个月的故事,安全方向值得看看。
Automated interface design shifts the bottleneck in computer-use agents from model quality to environment design, letting a small open model beat frontier systems on cost-adjusted performance. Builders of agent pipelines can apply the same trajectory-driven tool synthesis and population-based selection pattern to their own harnesses. The open-source release makes the method directly reproducible without proprietary APIs.
GPT-5.6 Sol 挖出一个 PHP 安全漏洞,OpenAI 说后面还有一批待公布,可以关注模型挖洞这条路。
在加拿大区域用 Bedrock 的朋友注意了,现在能直接调 GPT-5.6 到 GPT-6.1 这些模型,文章把权限配置和三种 API 调用方法都写清楚了。
Interface design itself becomes a learnable, evolving capability, meaning smaller open models can beat frontier systems on computer-use tasks at a fraction of the cost. Builders working on agents can reuse the open-source repository and apply the evolution loop to their own workflows. The results set a concrete cost-performance benchmark against GPT-5.5 and Claude Opus 4.8 that teams choosing an agent stack can weigh directly.
OpenAI 把 GPT-6 系列专门训练到能搞定 COBOL 这种老古董语言了,迁移银行系统代码准确率 97.9%,维护老系统的团队可以看看。
Meta 拿 AI 眼镜场景做了个新基准,GPT-5.3-Chat 有答案也才 38 分,问题出在自己挑时机说话上,方法部分挺有参考价值。
Legacy COBOL systems still run banking and government payroll worldwide, and a 97.9 percent migration accuracy claim suggests AI-assisted mainframe modernization is becoming practical rather than experimental. Teams maintaining COBOL or Fortran codebases should evaluate GPT-6.1 Sol against Anthropic's Claude and other coding assistants on their own migration test suites. The MultiLangBench scores also give enterprises a benchmark for comparing models beyond mainstream languages like Python.
A 94.1% score above the human noise ceiling suggests behavioural-simulation systems like Engram 1.0 can outperform frontier LLMs at modelling subgroups, not just answer questions. Product and research teams building personalisation, market research or user modelling tools now have a public baseline to test against. The 601 private cases make leaderboard gaming harder, so results should hold up better than fully public benchmarks.
Reproducibility is the missing layer in agent-driven science: Open-Science's checksummed versions and replayable .science packages turn a chat transcript into auditable evidence, which most agent frameworks do not attempt. Biomedical researchers and data scientists can attach their own data, pick any agent engine including local Ollama models, and share full analysis histories with collaborators. The Apache-2.0 licence and local-first design make it a concrete open alternative to closed cloud analysis tools, at no licence cost.
一篇讲怎么让 Agent 框架自动进化的论文,产出的 harness 比之前的 RHO 又快又便宜,代码已开源可以直接看。
PHREEQC 模拟结果固定不变,只是让 GPT-5.2 按需检索报告就掉 9.4 分,原因是它改用自己的输出文件瞎舍入。Claude 系列反而没受影响,做 agent 上下文管理的值得对比看看。
拿四个 LLM 模拟欧洲四国调查,看谁能复现真实的年龄差距,结果三个模型翻车,适合做问卷模拟的人看看。