Skip to content

GPT‑6 Cyber: Advancing the frontier of aligned cybersecurity agents

OpenAI

The cybersecurity model scores 48.5% on ExploitGym and is rolling out through the API and Codex to approved Daybreak Red customers.

Key points

  • Scored 48.5% on ExploitGym versus 32.1% for GPT-5.6 Cyber, and avoided out-of-scope services in 100% of alignment tests with its deployment prompt.
  • Earned a Critical cybersecurity classification under OpenAI's Preparedness Framework, up from High for GPT-5.6 Cyber.
  • Identified 81 novel high-severity Firefox vulnerabilities and generated working exploits for 948 Linux kernel vulnerabilities during internal research.
  • Introduces Daybreak Red for Critical Infrastructure, a specialized access group built with Dragos for defenders of operational technology and industrial control systems.

Why it matters

Authorized security researchers and defenders gain access to a frontier agent capable of autonomous exploit generation and patching while adhering to strict scope constraints.

Sources