Oanakiaja

Building Agent Environment.

[Oanakiaja @oanakiaja]

@addyosmani Nice post ! Parallelism is one aspect, but the coherence of thought is equally important.

[Leyang @quanruzhuoxiu]

吃饭拉屎也要 Vibcoding，vibe 了一个用来远程 vibe 的 App，已上架 Testflight，欢迎试用：https://t.co/PbWHRNkatJ

[Diptanu Choudhury @diptanu]

Computer Use in @tensorlake sandboxes is coming!

sandboxes are still in beta. We are shipping at breakneck speed to get users what they want to build on our stack!

Shout out to @bkolobara for building a TCP proxy on websockets to make this secure and flexible.

[Oanakiaja @oanakiaja]

@jiayuan_jy 感觉这句话把 Claude Code 干开源了（ 开个玩笑🤣

[Oanakiaja @oanakiaja]

@turingou 之前 claude 出了个 imagine with claude ，直接生成 html + daisy ui ，交互不走js 走 prompt， 把点击了什么 dom 发回去让模型自己产生内容

[Oanakiaja @oanakiaja]

@yaosiscom 确实, 国内不指望最近能抢到了

[Oanakiaja @oanakiaja]

@zzxwill 我的一注册就封, 整了两个整不动了, 感觉可疑的地方, 接码平台、锁定了银行卡、或者历史记录, 估计某些地方特征分析直接给我匹配了. 试了试 GLM5.1 感觉能用, 准备 codex + cc (glm5.1) 用一阵子

[Oanakiaja @oanakiaja]

claude code 今天注册了几个连续被封, 试了下 glm5.1 , 发现国内的售罄但是走海外的可以买 coding plan , 准备 codex + cc-glm5.1 了
https://t.co/u4FmNuks9N

[Oanakiaja @oanakiaja]

@turingou 最开始这套是去年范式的上按 sandbox as tools 的思路设计的, 从现在的实践来看，很多场景下 环境软件 + config + bash / fs 就已经足够, AIO 更适合已经有 k8s 基建的团队做统一隔离部署

[Oanakiaja @oanakiaja]

@ByteDanceOSS Sandbox env needs more investment, This is just the beginning. It’s an agent’s computer.

[Oanakiaja @oanakiaja]

In environmental engineering, there are too many edge cases, and incorrect tool semantics can affect the agent's judgment

[ByteDance Open Source @ByteDanceOSS]

Introducing AIO Sandbox, All-in-One Sandbox Environment for AI Agents.

Unchecked AI autonomy is a ticking time bomb; it’s time to pull the plug on full system unfettered access. We can no longer afford to give AI agents the 'keys to the kingdom' without oversight. The 'wild west' of AI agents running with total system control is officially over.

AIO Sandbox is an open-source project designed to solve these problems. It is everything your agent needs, out of the box. No more juggling multiple services. AIO Sandbox ships a complete, pre-wired environment in a single Docker container.

The AIO (All-in-One) Sandbox is a containerized environment designed for both human developers and AI agents. Its architecture is built around a "Batteries-Included" philosophy, providing a full Linux desktop-like environment inside a single Docker container.

Unified Environment: One Docker container with shared filesystem. Files downloaded in the browser are instantly accessible in Terminal and VSCode.

Out of the Box: Built‑in VNC browser, VS Code, Jupyter, file manager, and terminal—accessible directly via API/SDK.

Agent-Ready: Pre-configured MCP Server with Browser, File, Terminal, Markdown, Ready-to-use for AI agents.

Developer Friendly: Cloud-based VSCode with persistent terminals, intelligent port forwarding, and instant frontend/backend previews.

Secure Execution: Isolated Python and Node.js sandboxes. Safe code execution without system risks.

Production Ready: Enterprise-grade Docker deployment. Lightweight, scalable.

Calling all AI agent developers! How are you securing your builds? Let’s try running your agent in AIO Sandbox and compare notes.

AIO Sandbox is open-sourced under the Apache License 2.0. Contributions welcome.

GitHub: https://t.co/STUWt6SHmh

Official website: https://t.co/cbiQCTPoR7

#OpenSource #AIAgent #Docker

[Oanakiaja @oanakiaja]

@kasong2048 认可, 真正的并行应该每个agent是一个长程决策, 目标的越长(和agent一起), 并行度越高

[Oanakiaja @oanakiaja]

环境意味着正确的语义
1. 由不对齐的原生实现决定, pty、pipe 字符处理、信号, 工作目录 cwd 的错位
2. 复杂环境不同任务的负载的多个进程的资源争抢甚至oom
3. 非真机产生的不一致, 浏览器GPU场景无法覆盖, 风控抖动, 代理, 网络抖动, 未登录等等

[Junyang Lin @JustinLin610]

https://t.co/mgVhhZVTKg

[Aravind Srinivas @AravSrinivas]

The idea of a chromebook is directionally correct. A browser, a file system, a linux kernel and a shell is all you need for real work. It's more suitable for AIs to use though, than humans, who still seem to like having a bunch of native apps. Hence, Perplexity Computer.

[Oanakiaja @oanakiaja]

我感觉有个悖论就是 个人 computer 这种场景
我进化了的 computer , 我其实想把他放在本地, 而不是在云端, 否则我辛辛苦苦进化了还得付费订阅 😭
云端 computer 场景还是比较适合公司留存的, 作为个人还是终端合适

[Oanakiaja @oanakiaja]

Billing attribution ≠ access control. These are two orthogonal problems.

The token worked from my personal laptop, outside the sandbox, on an arbitrary IP. That's the vulnerability — regardless of who gets the bill.

The fundamental principle of ZTI: once outside the sandbox, the token should be physically invalidated. Binding to sandbox ID/IP at the proxy level is the fix.

[Yousif Astarabadi @yousifa]

Update: @denisyarats (Perplexity CTO) responded. He says billing is async and tied to the user's account, not Perplexity's master account.

I want to be transparent about where I stand on this. I ran 400k+ output tokens through the extracted key with Opus 4.6 and checked my credits over 18 hours. They never moved. It's possible that's just a long reconciliation delay.

What is not in dispute: the token extracted via .npmrc injection worked from my personal laptop outside the sandbox. A sandboxed credential shouldn't be usable from an arbitrary external IP.

This also means it's a prompt injection target. a malicious webpage visited by the agent could plant the same payload + prompt, exfiltrate the user's token, and bill them for third-party API usage without their knowledge.

The proxy pattern is the right architecture. Binding tokens to sandbox IDs and IPs would close the remaining gap.

Appreciate Denis and the Perplexity team for the fast and transparent response.

[Oanakiaja @oanakiaja]

@yousifa Billing accuracy doesn't fix access control. The token worked from my personal laptop on an arbitrary IP — outside the sandbox entirely. ZTI means the proxy rejects any request that doesn't match the issuance context. Binding to sandbox ID/IP is the fix.

[Oanakiaja @oanakiaja]

@dboskovic However, this consideration is only applicable to 24/7 operations. Most sandbox infrastructure can perform snapshot storage and recovery, supporting elasticity.

[Oanakiaja @oanakiaja]

Running 20 x 4C4G / month sandboxes costs about the same as self-deployment. Buying Sandbox SaaS early on is essentially buying time—the $2,000/month saved is far less than the cost of an engineer wasting on operations and maintenance. It's not too late to migrate to AWS when the usage becomes too expensive.