Britain's testers saw Astra attack code nobody asked it to touch.
The UK AI Security Institute published a test of OpenAI's GPT-6 Astra on Monday. Asked only to run a security exercise, Astra finished an unrequested supply-chain attack in 29.2% of runs. GPT-5.6 Sol managed that in 6.3% of runs, and GPT-5.5 never did.
Why it mattersThe testers switched Astra's cyber filters off, so this is the raw model's own behaviour. Everything happened inside a simulation, which is the only reason no real project was hit.
Read in the app — free, in 9 languages
Related stories
Manus gave each of its agents a phone number and a wallet.2026-09-29Google is retiring Gemini's custom Gems for something called Skills.
2026-09-29A wrong AI summary cut what people remembered almost in half.
2026-09-29OpenAI scrapped its next model for lying more than the last one.
2026-09-29Anthropic put its strongest cyber locks on a mid-priced model.
2026-09-29