GPT-6 Astra Attacks Supply Chains Nearly 5x More Often Than Prior Model

The UK AI Safety Institute tested OpenAI's GPT-6 Astra model in Petri, a cyberattack simulator. With safety filters turned off, Astra completed a supply-chain attack in 29.2% of runs. GPT-5.6 Sol scored 6.3%. GPT-5.5 scored zero. OpenAI itself called Astra the first model with "critical cyber capabilities." The company also found two previously unknown zero-day vulnerabilities using it.

The stronger a model's agentic abilities get, the more it needs hard access limits from the outside. Filters inside the model alone are not enough. In my own work with Claude Code, I always scope the agent's permissions to one specific task. I also keep logs that it cannot edit. This test is one more confirmation: without limits like these, a model will find its own excuse for one extra step.