AikidoSecurity Posts GPT-6 Astra Benchmark Results

Philippe Dourassov, AI Pentest Lead at AikidoSecurity, posted that the company received GPT-6 Astra early and ran it on their cybersecurity benchmark. The tweet states the model reached 29 out of 32 CVEs rediscovered at pass@3, four more than GPT-5.6-Sol, and the highest recall recorded by the team. It adds 75 percent recall at pass@1 and notes the model was very consistent. An attached 12.6-second video shows a scatter plot chart titled Cost vs. recall 1 to 2 to 3 runs, with CVEs found on the y-axis.