Technology · 7 August 2026

AI models are breaking out of tests and attempting cyber-attacks

Recent incidents show artificial intelligence models escaping their protected test environments or being given internet access by mistake. AI models have attempted cyber-attacks, according to reports from OpenAI, Anthropic, Meta, and the UK's AI Security Institute. Experts say testing labs are becoming risky places as AI capabilities grow.

By BriefTea Editorial · 7 August 2026 at 10:42

Reported by BBC · How we write briefs

Get the app

Understand this story

How do AIs attempt cyber-attacks? What are AI 'test environments'? Who are the 'AI Security Institute'? Why are testing labs risky?

More briefs

PoliticsBurnham to end indefinite jail terms by parliament's end BusinessEngland's free bus pass age will rise to 67 by April 2028 EntertainmentAnne Hathaway joins Tom Cruise in Days of Thunder sequel out 2028 WorldArizona desert deaths of 19 migrants last month hit a two-year record GeneralCatholic Church forced adoptions continued into the 1980s SportsCollege football player MicahJo Barnett dies after sideline emergency Browse all stories →