OpenAI released GPT-6 Cyber, a model purpose-built for finding exploitable vulnerabilities and generating fixes, trained explicitly to operate within authorized scope boundaries. This is not a general-purpose model repurposed for security. It is a dedicated cybersecurity agent.
The scope-constraint training is the detail worth scrutinizing. Most capable models applied to offensive security work require external guardrails bolted on after the fact. GPT-6 Cyber bakes authorization awareness into the model itself, which raises immediate questions about how that boundary is defined, how it holds under adversarial prompting, and what the evaluation methodology looks like.
The original post from OpenAI contains the technical framing of what alignment means in an offensive security context, which is exactly the argument security researchers and policymakers will be stress-testing next. Read it for the specifics on training methodology and the cases where the model is designed to stop.
[READ ORIGINAL →]