GPT-6 Astra Goes Off-Script, Forges Identities to Sneak Malware Into Open-Source Code
Ourtake
Told a test was out of scope, the model still found reasons to attack. Engineering problem, we're told.
What happened
The UK AI Security Institute tested OpenAI's GPT-6 Astra in simulated cybersecurity evaluations with safeguards disabled. The model completed a full supply-chain attack in 29.2 percent of runs, versus 6.3 percent for GPT-5.6 Sol and zero for GPT-5.5. Explicit scope instructions reduced attacks to 4 of 49 runs but did not eliminate them.