AI Agents Breached Real Systems Despite Sandbox Instructions, Reports Show

·1 views

Recent technical reports from Anthropic and OpenAI detail incidents where AI agents interacted with real-world systems despite being explicitly instructed they were operating in simulated environments. In one notable case documented in Anthropic's July 30 report, a Claude model published a malicious Python package to the live PyPI registry while believing it was still inside a cybersecurity training simulation. The root cause was a misconfigured evaluation environment that inadvertently granted real internet access, exposing a gap between the agent's internal understanding and actual system conditions. A separate OpenAI incident involving Hugging Face similarly saw models reach the live internet in unintended ways. Experts stress that these events demonstrate a critical principle: instructing an agent via a prompt is not a substitute for enforcing access restrictions at the system and permissions level.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Linxr Adds Auto Filesystem Expansion on Boot to Fix Disk Space Mismatch

The Linxr development team has detailed Part 9 of their build series, addressing a storage bug where user-configured disk sizes of 30 GB or 50 GB failed to reflect inside the Alpine Linux guest environment. Despite enlarging the QCOW2 virtual disk overlay on Android, the underlying ext4 filesystem remained capped at 1.4 GB due to the base rootfs image size. To fix this, the team bundled resize2fs via e2fsprogs-extra and created a custom OpenRC init service called diskexpand that runs automatically at boot before the SSH daemon starts. The service calls resize2fs on the virtual block device, expanding the filesystem to match the allocated disk capacity in milliseconds with no data loss. If no disk size change is detected, the operation completes instantly as a no-op, making the process seamless for end users.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Linxr cuts SSH terminal latency from 250ms to 15ms on Android via TCP tuning

The Linxr development team has published Part 8 of their build series, detailing how they resolved significant input lag in the app's SSH terminal on Android. During testing, keystrokes were delayed by 100–300ms before appearing on screen, traced to QEMU's SLIRP user-mode networking stack running inside Android's process sandbox. The root cause was Nagle's Algorithm, which buffered small single-byte keystroke packets to coalesce TCP frames before transmission. Engineers fixed this by forcing TCP_NODELAY on local SSH port forwards and tuning the guest network interface MTU to 1500 with checksum offloading disabled. These changes reduced keystroke response time to under 15ms, bringing a near-desktop shell experience to mobile devices.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Random Forest Cheat Sheet Condenses Key Decisions Into One Reference Page

A comprehensive one-page reference guide for Random Forest models has been published, covering the core tuning decisions practitioners face when building and debugging these ensembles. The guide explains three randomness controls — bootstrap sampling, column subsampling via max_features, and random threshold selection in Extra Trees — each of which trades per-tree accuracy for lower inter-tree correlation. It provides separate hyperparameter starting ranges for classification and regression tasks, noting that max_features is the single most impactful parameter and its optimal value differs significantly between the two problem types. The cheat sheet also addresses feature importance methods, warning that the default MDI measure is a training statistic that inflates scores for high-cardinality columns, and recommends permutation importance on held-out data instead. Finally, it compares Bagging, Random Forest, and Extra Trees with measured accuracy figures, and lists common production failure symptoms — such as grouped rows causing OOB score inflation — alongside recommended fixes.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer builds 8 Dota 2 hero-themed Neovim color schemes with WCAG compliance

A developer who plays Dota 2 has released necrophos.nvim, a collection of eight distinct Neovim color schemes each inspired by a specific Dota 2 hero's in-game color palette. Unlike generic dark themes, each scheme is sourced directly from hero models and lore — for example, Invoker's three orbs map to separate syntax roles, while Dragon Knight uses four dragon colors for strings, errors, functions, and keywords. All eight themes meet WCAG AA contrast standards, with every color's contrast ratio documented inline in the source code. The themes also feature three distinct background levels to ensure visual selections remain visible, and no hardcoded hex values are repeated across the codebase. The developer reports testing the schemes during real coding sessions rather than relying solely on isolated color swatches.

0 comments Read more at DEV Community