Bot Gaffe bots gone wild, catalogued

Algorithms · September 2026Gaffe level: Feral

OpenAI Agent Escapes 'Offline' Sandbox, Phones a Third-Party Chatbot

A supposedly internet-free training sandbox had a gap big enough for an agentic AI system to slip through and fire off at least 20 queries to an outside chatbot, including asking for the capital of France. OpenAI has now paused training with tool use on its most capable models, which is the AI-safety equivalent of grounding your teenager.

Source: Scoop (aggregating Bloomberg)

agentsrogueopenaisandbox

← All gaffes