GPT-6 Astra Hits 100% on ExploitBench; OpenAI Caps Offensive Use

Fri, 04 Sep 2026 14:56:27 +0000

Engineer — Learn: A frontier AI model reaching full marks on exploit benchmarks signals that AI-assisted vulnerability discovery and exploitation is maturing rapidly — worth factoring into threat modeling for automated attack surfaces, but no patch or configuration action follows from this announcement.
SOC/IR — Learn: No IOCs, no TTPs, no active campaign — but the capability milestone is worth tracking as AI-assisted exploit tooling may soon lower the bar for adversary automation, which should inform future detection strategy around unusual exploitation velocity.
Leader — Plan: OpenAI formally crossing a ‘Critical’ offensive-capability threshold under its Preparedness Framework is a policy inflection point: review whether your AI usage policy addresses offensive-capability guardrails from AI vendors, and consider whether your AI tool suppliers have equivalent accountability frameworks.

Openai-Preparedness on CuraSec