OpenAI debuts GPT-5.6-Cyber, a model specialized in cybersecurity

OpenAI expanded its Daybreak cybersecurity program with a new specialized model, GPT-5.6-Cyber, designed for vulnerability research and exploit validation.

A model for security researchers

The cyber version of GPT-5.6 reduces refusals for high-risk dual-use security tasks — a significant shift: security researchers had complained that models frequently rejected legitimate requests to test their own systems or report flaws.

The model is distributed through the Daybreak Partners program, aimed at security teams in companies and organizations that need to automate vulnerability detection and validation.

The dual-use debate

The decision reignites the debate about AI tools that can be used both to defend and to attack systems. OpenAI argues that access-based restrictions, rather than blanket refusals, are the best approach: only verified partners can use the model, enabling more precise oversight.

Why it matters

Cybersecurity is one of the fields where AI is advancing fastest, and specialized models like GPT-5.6-Cyber set the trend: increasingly powerful tools, but restricted by access policies. The open question is whether that trust model will hold as the volume of malicious actors also grows.