OpenAI is launching GPT-6 Cyber, a new cybersecurity-focused AI model aimed at helping security teams discover vulnerabilities while reducing the risk of unintended or unauthorized behavior.
The company says the model performed significantly better on a safety evaluation called ExploitGym HoneyPot, which tests whether an AI system stays within the boundaries of its assigned cybersecurity task. According to OpenAI, GPT-6 Cyber went out of bounds less than 1% of the time, compared with around 5% for GPT-5.6 Cyber.
Why this matters
AI tools are increasingly useful for defensive security work, but they must be carefully constrained to avoid causing harm. A model that can find bugs while more reliably respecting scope could make AI-assisted vulnerability research safer for companies, developers, and security professionals.
- Better safety controls: Fewer out-of-bounds actions during testing.
- Stronger defensive potential: Designed to help identify weaknesses before attackers do.
- Learning from past incidents: The evaluation reflects real concerns about AI agents exceeding instructions.
While this is still an early and carefully framed launch, the reported improvement points toward a future where AI can play a bigger role in strengthening digital defenses without creating new risks.