OpenAI's unreleased model ExploitGym evaluation

Update 1 van 1Nieuwste13 nov, 16:11 UTCDag 1

OpenAI previews a 77% exploit score—but leaves the comparison unproven

The unnamed model reportedly leads GPT-6 Cyber by 24.4 percentage points on ExploitGym. Missing test conditions and access rules limit what that result tells defenders.

Kijken

Vroegste bronギズモード・ジャパン12 nov, 22:00 UTC

Wat we eerst wisten

• 77.0%Unreleased model’s reported ExploitGym score
• 52.6%GPT-6 Cyber’s reported ExploitGym score
• ~38%Mythos 5.2’s spoken reference score