OpenAI Launches GPT-5.6-Cyber with Reduced Safeguards for Exploit Development

OpenAI unveiled GPT-5.6-Cyber, a cyber-permissive model built on GPT-5.6 Sol and available through a new "Daybreak Red" tier for authorised vulnerability research, exploit validation and security testing — with deliberately reduced refusals for dual-use cyber tasks.

Summary

GPT-5.6-Cyber is built on GPT-5.6 Sol and offered through a new "Daybreak Red" tier for authorised vulnerability research, exploit validation and security testing. Internal evaluations show it completes 95.0% of advanced-cybersecurity requests (exploit-chain development, authentication bypass, privilege escalation) versus 1.5% for the base model, and it outperforms its June predecessor GPT-5.5-Cyber (57.3%).

The release deliberately lowers refusals for dual-use cyber tasks, arriving days after OpenAI paused some internal activity following an Astra evaluation that found significant agentic-cyber capability gains. The digest links the launch to a live policy thread on AI-cyber permissiveness, with ACSC frontier-AI board guidance and NIST/FTC regulatory attention.

Date

• Announced: 2026-08-11

Source

• The Hacker News — 2026-08-11