OpenAI introduces GPT-6 Cyber for aligned cybersecurity agents

Summary

OpenAI has introduced GPT-6 Cyber, a model for security research that is better at finding and fixing vulnerabilities and more reliable at staying within researcher-set boundaries. On the company's own evaluations it scores 48.5% on ExploitGym, a 51% relative improvement over GPT-5.6 Cyber, and avoided exploiting out-of-scope services in 99.93% of alignment tests without deployment prompts (100% with them). OpenAI says the model produced 948 working Linux kernel privilege-escalation exploits and uncovered 81 previously unknown high-severity Firefox vulnerabilities, some of which Mozilla has already patched. Access is initially limited to approved Daybreak Red customers rather than general availability.

Why it matters for our work

The range of repetitive discovery and validation work that security teams can hand to agents is visibly widening, while human researchers shift toward review and authorization. The performance figures are OpenAI's own evaluations, so real-world verification remains each team's job.

Translated from the Korean original. Summaries may be translated and edited. Commentary reflects our perspective; forecasts remain the source’s views.