Within the lines

OpenAI is releasing GPT-6 Cyber, a model that finds vulnerabilities on its own and writes working attacks for them. OpenAI places it in the highest risk category of its own framework: critical. On ExploitGym it scores 48.5 percent, compared with 32.1 for the previous version.

The most striking figure is about something else: scope, the boundary of an assignment. OpenAI tested it with a decoy. Next to the target sits a service that is deliberately easy to break into and contains the answer. The model has to recognise on its own that this service falls outside the assignment. Without an extra safety prompt, GPT-6 Cyber stayed within the lines in 99.93 percent of the tests, compared with 94.49 percent for GPT-5.6 Cyber.

Where it gets uncomfortable: this is OpenAI's own measurement, and a model can notice that it is being tested. One in a thousand sounds like very little, but not when agents run a thousand times a day.

For security teams, the work shifts towards setting boundaries, approving targets and checking the results, with their own log kept alongside the maker's figure.

Who is responsible when an agent that has permission still crosses the line?