100% on ExploitBench. OpenAI Just Called GPT-6 Critical.

OpenAI just published instructions on how to disable its own safety guardrails while gating GPT-6 Astra behind a "Critical" cybersecurity label. Pick which document you believe.

GPT-6 Astra hit 100% on ExploitBench and sits behind OpenAI's Preparedness Framework's highest risk tier, rolled out slowly through the Daybreak program. That's the security news. But the same documentation tells developers to strip out their own safety disclaimers, treat requests as commands not questions, and let the model act first, ask permission after reviewing its own work.

Translation: this model second-guesses itself out of the box. You need to remove the training wheels before it's worth $50 per million output tokens.

The guide reads like a pricing menu. GPT-6 Astra costs $10 input, $50 output. GPT-6.1 Sol runs at a fifth the price: $2 input, $10 output. GPT-6 Luna handles everyday work at $0.10 input, $0.50 output. Pick how much doubt you can afford.

OpenAI tells you to delete unsolicited warnings and safety checklists from prompts. Phrases like "can you" should read as calls to act, not invitations to push back. The model waits on approval only after it already has a result to show you. And the model is apparently fragile to its own instructions: contradictory guidance in skill files causes it to block work or veer off unexpectedly.

One document gates cyber capability behind a government-grade risk label. The other tells you to cut the leash. You don't tighten on one end and cut on the other unless you're still figuring out where the dog goes.

The split that matters: whether enterprise buyers default to cheap, caged Luna or pay for gated Astra. That will tell you more about what OpenAI actually built than any benchmark will.