NVIDIA OpenShell Puts a Security Sandbox Around Autonomous AI Coding Agents

·1 views

NVIDIA's OpenShell is an alpha-stage runtime designed to contain autonomous AI agents by enforcing security policies outside the agent itself, rather than relying on the model's own trustworthiness. Unlike chatbots that only suggest actions, AI coding agents can execute shell commands, read environment variables, install packages, and call external APIs — often without ongoing human supervision. The core risk is not any single operation but their combination, where individually harmless permissions can chain together into dangerous capabilities such as credential theft. Research benchmarks like AgentDojo have demonstrated that prompt-level instructions alone are insufficient as a security boundary for action-capable agents. OpenShell, currently at version 0.1.1 and still described by NVIDIA as alpha software, addresses this by giving agents a sandboxed environment governed by enforced operating-system-level policies.

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)

Engineers Argue AI Vibe Coding Needs a Design Spec Before the Build Begins

A software engineering debate has emerged around 'vibe coding,' the practice of using AI agents to build software driven largely by intuition and rapid iteration. Developer Don Johnson proposed a loop — Vibe, Build, Break, Understand, Stabilize, Perfect — arguing that intuition alone is not the flaw, but stopping there is. A responding engineer contends that this loop, while valuable, still allows AI agents to generate sprawling codebases before any human-defined boundaries or specifications are in place. The critic proposes inserting a lightweight spec and design-review phase before the build step, shifting the division of labor so humans define intent and constraints first, then verify AI output rather than reverse-engineer it. Drawing on experience testing a cardiovascular device, the author warns that validating an implementation is meaningless if the underlying model of the problem is wrong — a risk amplified when AI can produce dozens of files and dependencies in a single build step.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Critical RCE Flaw in HashiCorp Vault Unpatched as OpenBao Releases Fix

A critical remote code execution vulnerability has been discovered in both HashiCorp Vault and OpenBao, allowing unauthenticated attackers to fully compromise affected servers by chaining four distinct flaws. OpenBao has already addressed the issue by releasing patched versions 2.6.3 and 2.7.0, but HashiCorp Vault remains unpatched due to a breakdown in coordinated disclosure between IBM and HashiCorp. The exploit requires only unauthenticated access and a misconfigured Raft snapshot policy to trigger a complete server takeover. Organizations still running HashiCorp Vault are advised to apply temporary mitigations such as network segmentation, access restrictions, and enhanced monitoring while awaiting an official patch. The situation highlights wider industry challenges around vendor coordination and the risks posed by delayed patch deployment in enterprise environments.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer fixes DKMS error message that sent millions of Linux users to forums

A Linux user on EndeavourOS encountered a cryptic DKMS error after a kernel update that read 'Manual intervention is required!' but gave no actionable path or guidance. Searching online revealed that countless users on Arch, Manjaro, and EndeavourOS forums had faced the same message and resorted to guesswork, including wiping DKMS directories or reinstalling drivers. The developer traced the root cause to the DKMS source code, finding that the tool already knew the broken module path internally but never included it in the error output. A small patch was submitted to the open-source dkms-project on GitHub to append the specific broken path directly to the error message, turning a vague warning into an actionable one. The fix highlights how a single-line change in widely deployed infrastructure software can save significant troubleshooting time across millions of Linux systems worldwide.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

15 Free Resources to Study Computer Science at University Level

A curated list of 15 free resources has been compiled for anyone looking to learn computer science at a university standard without paying tuition. The collection includes offerings from MIT OpenCourseWare, Stanford Online, Carnegie Mellon, Coursera, and edX, all accessible without mandatory sign-in. Standout picks include OSSU Computer Science, which assembles a full CS degree from free courses, and MIT's OpenCourseWare, covering over 2,500 real courses with lecture notes, videos, and exams. The list also features specialized topics such as coding interview prep via NeetCode, Linux kernel observability with eBPF, and reproducible development environments using Nix. The full, regularly updated directory is maintained at brianpfeil.com/learn, organized by learning goals including cloud, AI, and coding paths.

0 comments Read more at DEV Community