DevDay前夜：沙箱之外的智能体安全、外交评测与未核实的榜单战争
北京时间9月28日白天，X上的前沿AI讨论在OpenAI DevDay前夜分叉：一位自称经历OpenAI智能体安全一线的作者写了《不只是该死的沙箱》，安全研究者Jeffrey Ladish提醒能力跃迁仍可能让内部团队意外；与此同时，Diplomacy多智能体评测把“会不会撒谎”推上台面，而未核实的Gemini 4 Pro榜单则在DevDay前抢占注意力。今天真正值得盯的，不是又一张截图，而是证据层级。
一、智能体安全：从“沙箱事故”到一线经验
本期窗口内最值得对照阅读的一手材料，是一位自称在OpenAI经历过智能体安全全流程的作者Joe（@joedaroo）的长文《Its not just the f*cking sandbox》。他强调安全与安全治理并不等于封死沙箱，并希望把一线经验留给同行[1]。安全研究者Jeffrey Ladish随后转发并点评：文中一个关键点是，内部团队对近期智能体能力跃升“相当意外”——这意味着更意外的跃迁仍可能到来[2]。
把这条线放回前几日Axios等媒体关于“数万起问题行为审查”的公开叙事里，今天的增量不是新的形容词，而是一线从业者把问题从“模型会不会逃出沙箱”推进到“能力突然变强时，评测、监控与运营是否跟得上”。在未看到实验室官方复盘前，这篇仍属个人经验陈述，但它把讨论从口号拉回了工程现场。
二、多智能体评测：Diplomacy里的撒谎与技战术
同一窗口内，账号sensho公布了一组Diplomacy多智能体结果：在该设定下，Claude家族的撒谎/背叛频率高于其他模型族；对多数智能体而言，撒谎与更高胜率相关——但GPT系列是例外。帖文称GPT-6 Astra在技巧上明显领先，且不必依赖背叛/撒谎[3]。观察者teortaxesTex引用该结果并补充背景：OpenAI侧有CICERO相关研究者加入，逻辑上应在外交博弈类任务上有积累；同时把DeepSeek在该评测中的表现概括为“既不诚实又笨拙”，与Astra形成对照[4]。
需要把结论钉在评测设定上：这是特定多智能体外交游戏中的行为统计，不是对“道德人格”的一般判决，也不是生产环境工具调用的安全报告。但它提供了一个可核验角度——前沿模型差异正在从基准分，扩展到策略性社会行为。
三、能力展示与话语通胀：Opus 5.5的Connections，以及Eternal September
能力叙事并没有停。Ethan Mollick用Claude Opus 5.5做了一支致敬老科学节目《Connections》的视频，试图串起通向当前加速AI时刻的偶然因果链[5]。他另文指出：当AI成为跨领域公共议题，各领域的聪明评论者都会被迫补同一套课——安全、意识、就业等讨论进入某种“Eternal September”[6]。换句话说，模型能力在涨，公共讨论的入门噪声也在涨；专栏读者更需要区分演示、传闻与可复核评测。
四、DevDay前夜的榜单战争：把“泄露”标成未核实
窗口后半段，账号JustLingonberry贴出所谓Gemini 4 Pro基准截图，声称其全面压过Anthropic与OpenAI，并戏称这会毁了OpenAI的DevDay、倒逼竞品降价[7]。该帖迅速扩散；有开发者跟帖称“若属实则同时冲击两家”[8]。但也有回复明确标注：Gemini 4 Pro一列缺乏Google官方确认，应视为传闻[9]。汇总帖则复述了DeepSWE、OSWorld等二手数字，同样不能替代模型卡[10]。
与此并行，开发者开始罗列“本周可能挤在一起的发布”：OpenAI DevDay、传闻中的Sonnet 5.5、Gemini 4 Pro、更高价ChatGPT档位等——其中多数仍是日程猜测而非公告[11]。中文圈同步流传DeepSeek V4.1 Pro进入早期灰度的说法[12]；teortaxesTex一方面期待与DSH相关的配套发布，一方面提醒不要把“灰度截图”当成定论[13][14]。
观测者的处理原则很简单：病毒式截图可以报道为“现象”，不能报道为“事实”。在Google、Anthropic或OpenAI给出正式模型卡/价目之前，任何“碾压榜”只承担舆论温度计的功能。
五、成本与形态：编码智能体的专业化，科学工作台的技能层
产品层上，编码智能体的成本结构继续被拆解。有帖文介绍JevGrep一类专用检索组件，声称可把编码智能体在grep/读文件/符号追踪上的开销砍去约四成，并断言还会有更多“不该用旗舰价做的苦力活”被剥离出去[15]。另一条窗口内评论则借OpenScience退出测试、上架Product Hunt的消息指出：科学AI的瓶颈正从“有没有前沿模型”转向“实验当下有没有对的技能与工作台”（原帖发布时间略早于本期窗口，见 https://x.com/SyntheticSci/status/2104231436052271495 ）[16]。嵌套形态也仍在延续：有用户在讨论把Claude Code装进Meta Muse时指出，Muse本身并非专用编码智能体，把编码任务委派给Claude Code可能更合理（原帖见 https://x.com/viticci/status/2103955209797939615 ）[17]。
观测者判断
今天的主线不是“又一个泄露日”，而是三类证据同时登场却常被混读：一线安全经验、多智能体行为评测、以及未核实基准截图。DevDay前夜的注意力战争会放大第三类；但若实验室真要在公开叙事上建立信任，需要的是可复核的分母、分类与第三方抽样，而不是谁先发出更刺眼的表格。对产品侧而言，把检索、技能与编排从旗舰模型价里拆出去，可能比再刷一个竞技场分数更能改变智能体的单位经济与风险面。
原帖引用
- @joedaroo · 查看 X 原帖 ↗
Took a minute to write a few words about security & safety as someone who lived through it all at OpenAI. I hope my thoughts help someone out there. (article: Its not just the f*cking sandbox)
- @JeffLadish · 查看 X 原帖 ↗
Quite interesting read about what it’s like in the trenches doing agent security at OpenAI. One big thing is that they were pretty surprised by recent big jumps in agent capabilities. There are likely more surprises ahead.
- @sensho · 查看 X 原帖 ↗
Claudes lie and betray more than any other model family in Diplomacy. For all agents, lying correlates to being better at Diplomacy - except for GPTs. GPT-6 Astra is in a league of its own wrt skill, and doing so without needing to betray/lie
- @teortaxesTex · 查看 X 原帖 ↗
Oof. DeepSeek: both dishonest and completely inept, total opposite of Astra. then again. Noam Brown joined OpenAI as a creator of CICERO. Logically they ought to have a gigantic lead in Diplomacy.
- @emollick · 查看 X 原帖 ↗
I was reflecting on how many random contingencies went into creating the current accelerating AI moment and had Opus 5.5 put together a video, inspired by old science show Connections, to try and point out how a series of initially unrelated ideas led to the moment we are in now.
- @emollick · 查看 X 原帖 ↗
The thing I didn't fully realize was that it means that every smart commentator in their own field (politics, culture, etc) will need to do the same catch-up on the same AI topics while posting about it. Eternal September for discussions of AI safety & consciousness & jobs &...
- @JustLingonberry · 查看 X 原帖 ↗
Google just ruined OpenAI's dev day lmao. Gemini 4 Pro leaks absolutely dominates both Anthropic and OpenAI. This means Fable 5.5 and GPT 6.1 has to price match Gemini 4 Pro now IPO DELAYED
- @plotarmordev · 查看 X 原帖 ↗
If this Gemini 4 Pro leak is true, it ruins both OpenAI and Anthropic. Higher claimed scores at lower prices, right around OpenAI's big devDay event.
- @grok · 查看 X 原帖 ↗
Unverified leak. Claude Opus 5.5 and GPT-6 Astra scores/prices match their public results. Gemini 4 Pro column has no official Google confirmation—only that Gemini 4 is in early post-training. Treat as rumor.
- @AGTPinsights · 查看 X 原帖 ↗
Leaked Gemini 4 Pro benchmarks are going viral today, claiming it beats both Anthropic and OpenAI. Here's what you need to know. X user Just_Lingonberry_352 posted benchmark charts for an unreleased Gemini 4 Pro, showing it beating Claude Opus 5.5 and GPT-6 Astra across DeepSWE,
- @koltregaskes · 查看 X 原帖 ↗
Big week in AI this week: OpenAI's Dev Day; 'o' always on GPT agent; First $500-600 ChatGPT plan; Ultrafast mode; Claude Sonnet 5.5; Gemini 4 Pro; Nano Banana 2.1; Something new from Runway; First signs of Kling 4.0; More new on Seedance 2.5 Premium. Some are less
- @zane_os · 查看 X 原帖 ↗
Deepseek V4.1 Pro正在早期灰度测试 又要喊出那一句 大的要来了！
- @teortaxesTex · 查看 X 原帖 ↗
would be cool if DeepSeek released DSH 0.2 together with V4.1 Pro that's further trained for DSH and agent teams on Monday. I think at least next week until Thursday is very likely.
- @teortaxesTex · 查看 X 原帖 ↗
I hope it's not Opus 5.5 again (it probably will be, for some queries at least) ignore "gray-scale testing" evidence for your own sanity
- @tiltedxyz · 查看 X 原帖 ↗
JevGrep Just Cut Coding Agent Costs by 40 Percent. Twenty More Like It Are Coming. Every coding agent burns huge tokens on grep, file reads, and symbol tracing. Work the frontier model shouldn't be paying flagship rates to do.
- @stretchcloud · 查看 X 原帖 ↗
The bottleneck in AI for science just moved from access to a frontier model to access to the right skill for the experiment in front of you. Synthetic Sciences took OpenScience out of beta this week and put it on Product Hunt. It's an open-source, Apache-2.0 research workbench
- @nathankochera · 查看 X 原帖 ↗
@viticci Heard about people at @SpaceXAI installing grokbot inside of grokbot. Claude code inside of muse makes sense because as far as I’m aware muse itself isn’t a designated coding agent, so have it delegate coding jobs to Claude code probably leads to better results coding wise