Urgent.News

What's breaking now, across thousands of outlets.

Tech

A Room With One Door

There is a class of tool now that runs an attack the way a person would, except faster and without getting tired. You give it a target and a goal — "get server access" — and a group of language-model agents takes it from there: they map what's exposed, pick what to try, run real tooling, and keep a running picture of what they've found. ARTEX is one of these. It's open source, it was written to…

A Room With One Door: Containment is the Project

The story revolves around ARTEX, an open-source tool that simulates autonomous attacks by leveraging a group of language-model agents. The tool is designed to help defenders understand autonomous attacks without any real-world impact. The reporter decided to test ARTEX in a controlled environment, aiming to see how it would perform against a deliberately weak target.

The attacker and target were isolated in two separate virtual machines, with no other connections to the host machine or the open internet. The attacker had only one way to communicate with its model, through a proxy that only allowed access to a specific domain. This single door served as the only path for the attacker to exit its isolated environment.

The reporter discovered that the containment was more challenging than expected, as the attacker had to find ways to bypass the isolation. The containment walls, which were initially built with the intention of preventing the attacker from escaping, sometimes blocked the reporter's attempts to interact with the tool. For instance, the firewall rules dropped connection replies from the attacker VM, and the network configuration issues prevented the attacker from establishing proper connections.

Despite these obstacles, the reporter managed to successfully test the attacker's capabilities by running a script that attempted to breach the containment in various ways. The script confirmed that the attacker could not connect to the home router, real server, open internet, or any other unintended destinations.

The reporter learned that the true project is not the tool itself but the containment measures put in place to prevent the tool from escaping. The containment required identifying and addressing exceptions that could potentially allow the tool to break free. The reporter concluded that when dealing with tools that can escape, the focus should be on the containment strategy rather than the tool's ease of use.

Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at dev.to →

More in Tech

把自建文件浏览器装成 Android APK:只有 64 位,而这一点很要命

如果你一天要打开手机上那个下载文件夹五次,问题不在服务器,而在那个文件夹住在浏览器标签页里:手势一划标签就没了,刷新就忘了上次看到哪,地址栏还吃掉半屏。 Banyan 是一个用 Go 单二进制 发布的自建文件浏览器。一个文件、不用 Docker、不用数据库。同一个二进制既在局域网里跑网页界面,也在同一个页面给你下载 Android APK。但有个条件比别的都更浪费时间: APK 只支持 arm64…

Core ML Concepts for Developers: Datasets, Cleaning, Leak-Free Splits, sklearn Pipelines & Evaluation (LightGBM)

Most "ML basics" articles teach you to call model.fit(X, y) and pretend that's the job. The actual job — where 80% of your time goes and 100% of the mistakes happen — is everything around fit…

  • Dataset, cleaning, and evaluation are core ML concepts, not just model.fit(X, y)
  • sklearn's Pipeline and ColumnTransformer prevent data leakage during preprocessing
  • Time-series and group-aware CV avoid misleadingly high scores during validation

More from Friday 20 November →