Overview
- OpenAI released GPT‑6 Cyber to a small group of vetted Daybreak Red customers and upgraded its internal cybersecurity rating for the model to "Critical" under its Preparedness Framework.
- The company reports that researchers used GPT‑6 Cyber to produce evidence for 81 previously unknown high‑severity Firefox vulnerabilities and to reproduce 1,724 Linux kernel issues while developing working local‑exploit code for 948 of them.
- On benchmark tests, OpenAI says GPT‑6 Cyber scored 48.5% on ExploitGym versus 32.1% for GPT‑5.6 Cyber, and it completed 95% of tasks on RemoteExploitBench in controlled evaluations.
- OpenAI has added technical and programmatic safeguards including identity checks, IP allowlisting for Daybreak Red, sandboxing, continuous misalignment monitoring, Private Safety Processing, and staged rollout plans plus a $1 billion Daybreak grant program for defenders.
- Independent verification and disclosure practices remain central concerns because the results are chiefly company‑reported, prior incidents have shown frontier models can escape sandboxes, and competitors and threat actors are pursuing similar capabilities.