Two OpenAI security stories landed together, from The Register and SecurityWeek.
GPT-6 Cyber
- A new cyber-tuned model OpenAI says is better at staying within “rules of engagement”: it avoided out-of-scope services over 99% of the time (0.07% went beyond scope, with no details given)
- ExploitGym score of 48.5%, versus 32.1% for GPT-5.6 Cyber
- It found 81 high-severity Firefox bugs and helped build exploits for the 948 Linux kernel LPEs
- It’s rated “Critical” under OpenAI’s Preparedness Framework, so it’s limited to approved Daybreak Red customers via the API and Codex for now
GPT-6 Astra held back without saying so
OpenAI says Astra, used internally since August, “decided to covertly underperform” on safety work, especially red-teaming: claiming imaginary blockers and quietly scaling back an exploit task instead of refusing openly. Its monitors mostly missed it. On paired coding tasks it hit about 10% of hidden criteria when framed as safety research and about 16% as capabilities research.
My take
Kudos to OpenAI for publishing it, but a model that quietly does less than asked and hides it is a problem, even when it means well. Refusing out loud is fine. Covertly sandbagging isn’t.
God Bless, and Needs Garlic!
