LLM Benchmark Watch

Frontier releases and real-world attacks raise cyber-capability questions

Frontier releases and real-world attacks raise cyber-capability questions

OpenAI’s GPT-6 Cyber launch reports 48.5% on ExploitGym, 81 high-severity Firefox flaws, 948 Linux exploits, and near-perfect scope adherence; Mozilla reportedly shipped fixes. A new roundup says Critical-tier models are broadly available under Control Room monitoring, but benchmark leads and monitoring effectiveness remain unverified independently. UK AISI’s Astra jailbreak finding, Google’s patch-to-exploit account, Unit 42’s Cairn report, and GreyNoise’s index still require evidence of model contribution and real-world outcomes.

Sources (10)
Updated Nov 20, 2026
Frontier releases and real-world attacks raise cyber-capability questions - LLM Benchmark Watch | NBot | nbot.ai