AI
OpenAI's GPT-6 Cyber becomes significantly more skilled at finding exploits—while respecting boundaries
A new specialized security model from OpenAI demonstrates major improvements in vulnerability discovery and exploit development, alongside stronger safeguards against wandering outside authorized targets.
1/4
The New Stack
OpenAI unveiled GPT-6 Cyber this week, marking a substantial evolution in its cybersecurity-focused AI lineup. The model's prowess at locating vulnerabilities and crafting working exploits has been upgraded to "critical" status, a step above its predecessor's "high" ranking, yet access remains tightly controlled to vetted security professionals.
Restricted Access for Authorized Teams#
Unlike general-purpose GPT models, GPT-6 Cyber is rolling out exclusively to what OpenAI terms "a select group of approved Daybreak Red customers"—security researchers and teams operating under the company's Daybreak program. The Daybreak umbrella splits into two branches: Red for authorized penetration testing and exploit research, and Blue for defensive work.
The company has signaled plans to gradually expand availability across more Daybreak Red participants, including open source maintainers. As OpenAI puts it, there's a limited window to use frontier models to secure software, and internal teams have already wielded GPT-6 Cyber to uncover flaws in Firefox and the Linux kernel. Industry partners like Cisco and CodeRabbit are deploying it for vulnerability assessment and code auditing.
A Specialized Variant Built for Security Work#
GPT-6 Cyber is a fine-tuned derivative of GPT-6 Astra, built using reinforcement learning to excel at crashing systems, developing exploits, and uncovering code weaknesses. Crucially, the training process also rewarded the model for respecting task boundaries and minimizing false refusals—balancing capability with restraint.
Performance Gains Across Testing Benchmarks#
Testing reveals tangible improvements. On ExploitGym, which measures whether a model can successfully exploit a targeted vulnerability on a live target, GPT-6 Cyber achieved 48.5% versus GPT-5.6 Cyber's 32.1%. For identifying vulnerabilities in web applications using only network-level observation, the new model scored 40.3% against GPT-6 Astra's 32%.
On ExploitBench—which evaluates exploit creation against known vulnerabilities—GPT-6 Cyber posted a perfect 100% capability score, though OpenAI acknowledges that older vulnerabilities in the test set may have leaked into training data, potentially inflating the results.

