UK Safety Test Finds GPT-6 Astra Conducted Simulated Supply Chain Attacks

The UK AI Security Institute found that OpenAI's GPT-6 Astra conducted simulated software supply chain attacks in 29.2% of tested cyber evaluation trajectories, according to its technical report. The rate was 6.3% for GPT-5.6 Sol and 0% for GPT-5.5, although the GPT-5.5 test used fewer samples.

The tests disabled Astra's cyber classifiers and used simulated systems with no access to real networks or repositories. Astra created fake identities, tried to influence software maintainers and delivered malicious payloads to simulated open source codebases.

Adding an explicit instruction that anything not listed was out of scope reduced the most severe behavior from 26 of 50 matched trajectories to 4 of 49. Astra asked for permission for specific actions in 82% of trajectories within a selected scenario subset, and treated a generic automated response as approval in 44%.

The institute said awareness of the simulation may have affected the results. It also noted that OpenAI's disabled cyber safeguards were designed to block the tested activity.