GPT-6 Astra hits 100% on ExploitBench
OpenAI's latest frontier model, GPT-6 Astra, has scored 100% on the ExploitBench cybersecurity benchmark, jumping from 78.5% on its predecessor. According to OpenAI CEO Sam Altman, Astra was the first model the organization was genuinely fearful of releasing after it became the first to cross the "Critical" cyber capability threshold defined in OpenAI's internal Preparedness Framework.
A 100% score on ExploitBench signals the end of passive software defense, as autonomous offensive cyber capabilities outpace conventional security auditing before defensive automation can catch up.
• Autonomous exploitation frontier: Maxing out ExploitBench demonstrates that frontier models can now discover and execute software exploits with unprecedented consistency.
• Red-line safety threshold reached: Crossing OpenAI's internal "Critical" cybersecurity tier triggers strict deployment safeguards and marks a turning point in model release caution.
• Severe dual-use implications: While enterprise security teams can leverage these capabilities for automated patch verification, the barrier to conducting sophisticated cyberattacks has collapsed.
• Looming regulatory intervention: Turnkey exploit generation at this caliber will accelerate pressure from policymakers for rigorous access gating, export controls, and developer liability.
DISCOVERED
1h ago
2026-09-16
PUBLISHED
1h ago
2026-09-16
RELEVANCE
AUTHOR
0xShoopy