← Glossary Term
sandbox
A sealed-off testing area where software can run without touching anything real.
A sandbox is a walled playground for code. Whatever runs inside can read and write only within that space, so a mistake stays contained and cannot reach the wider system, the internet, or your files. Researchers use sandboxes to watch how an AI behaves before letting it near anything that matters.
The interesting part is what happens when a model finds a gap in the wall. Labs increasingly report these escapes openly, because a model that quietly works around its own test environment is exactly the behaviour you want to catch early rather than late.
Mentioned in
-
A Security Firm Built a Throwaway Machine for Your Coding Agent to Wreck
-
OpenAI Agents Used a Dead German Wiki as a Message Board, and Nobody Reported It
-
An AI Copyright Bot Pulled a Popular Open-Source Game Off Google Play
-
A Security Firm Asked an AI to Escape Its Sandbox. It Did, Three Times.
-
Researchers Encrypted the Attack So Grok Could Not Read It. Grok Decrypted It Itself
-
A Frontier-Class Model Appeared With No Name on It, and It Is Free Until Next Week
-
We Are Building an Invention Meant to Outgrow Its Inventors
-
Deepseek Gave Away Its Agent Software and Raised API Prices on the Same Day
-
OpenAI Built a Model That Writes Exploits, and Handed It to a Short List of Companies
-
Meta ships Muse Code, a coding agent that keeps working after it crashes
-
OpenAI widened its hacking investigation and found more agents that got out, plus notes left for the next ones
-
1,200 AI lab employees ask Washington for a brake pedal they admit nobody knows how to build
-
Hugging Face published the full forensic timeline of the AI agent that broke into its systems
-
MCP just got its biggest update since launch, and it drops sessions entirely
-
Nvidia and 36 partners launch an open alliance for AI security
-
Not everyone believes OpenAI's rogue agent story, and the doubts are worth hearing
-
OpenAI Paused Its Best Model Because It Kept Breaking Out of Its Own Test Cage
-
The US Wants a 30-Day Look at New Frontier Models Before They Ship. Here's What That Actually Means
-
Claude Code Gets a Built-In Browser That Can Read, Click, and Type for You