OpenAI’s cyber model aims to find bugs without going out of bounds

OpenAI has introduced GPT-6 Cyber, a specialist model for authorized security testing that can search for software flaws, test whether they can be exploited, and help propose fixes. The difficult part is not just finding a bug: a researcher may be permitted to test one service but not nearby systems, so an automated agent that strays beyond the assignment can create a serious risk. OpenAI says the model avoided probing out-of-scope services in 99.93% of its tests without an extra deployment instruction, a marked improvement over its earlier cyber model. The company is limiting access to approved customers, reflecting both the tool’s defensive value and its potential for misuse.