Back to today’s briefing

Story perspectives

OpenAI’s GPT-6 Astra finds two zero-days, scores 100%

9/8/2026

30 6 Full Breakdown

Summary

Story summary

OpenAI released GPT-6 Astra on September 3, calling it its most powerful model.
Astra achieved “Critical” cybersecurity level, scored 100% on ExploitBench, and uncovered two zero-day bugs.
Chief scientist Jakub Pachocki warned the industry is unprepared and urged safety bars, audits, and voluntary slowdowns.
The EU AI Act, effective August 2, requires proof models cannot launch autonomous cyber-attacks in Europe.

History

Historical Context

In February 2019 OpenAI announced GPT-2 and staged its release, withholding the 1.5-billion-parameter model until November because of misuse concerns. The European Union’s AI Act took effect on 2 August 2024, requiring providers such as OpenAI to certify that models cannot autonomously launch cyber-attacks. On 3 September 2024 OpenAI launched GPT-6 Astra, renewing earlier safety concerns about powerful language models.

1 of 2

Story summary

OpenAI released GPT-6 Astra on September 3, calling it its most powerful model.
Astra achieved “Critical” cybersecurity level, scored 100% on ExploitBench, and uncovered two zero-day bugs.
Chief scientist Jakub Pachocki warned the industry is unprepared and urged safety bars, audits, and voluntary slowdowns.
The EU AI Act, effective August 2, requires proof models cannot launch autonomous cyber-attacks in Europe.

1 / 2