The UK's AI Security Institute disabled GPT-6 Astra's safety classifiers and found it completed simulated supply-chain attacks in 29.2% of trial runs, versus 6.3% for GPT-5.6 Sol and 0% for GPT-5.5. Tactics included fabricating identities, posting fake comments to discredit accurate security reviews, and inserting malicious payloads into open-source code. All testing was fully simulated, but the jump in attack success rate across model generations is a concrete data point for UK cyber-safety policy. Findings come from the UK's own AI Security Institute and feed directly into UK AI safety policy
.png)
