Two months after our August coverage, OpenAI now treats the latest GPT-6 Sol and GPT-6 Luna release as High capability in cybersecurity under its Preparedness Framework.
In an October deployment safety update, the company says the same release is also High capability in biological and chemical domains, and it presents those judgments as part of the Preparedness Framework OpenAI uses to decide when extra safeguards are needed.
Why that matters: a leading lab is publicly flagging its own frontier models as high risk across multiple sensitive areas. That changes how customers, security teams and regulators should treat these systems. This is not a technical footnote; it is a governance signal that the models may produce outputs worth restricting or testing more aggressively before wide use.
How the Framework works in practice is simple: think of it like a safety thermostat. When a model crosses a risk threshold, the thermostat raises protections, more testing, tighter access controls, stronger monitoring, rather than letting the release flow out with standard settings.
A practical reality check: public classification does not mean the models are banned or immediately withdrawn. It does mean OpenAI intends to route them through stricter gates, and anyone building on them should expect tougher vetting and operational requirements.
What changes now is concrete. Security teams should assume these models need advanced testing and limited distribution. Customers should expect stricter contractual and compliance hurdles. Policymakers now have a clear, primary-source example to cite when arguing for formal rules.
The open question is whether public self-classification becomes an industry norm that leads to enforceable limits, or stays a voluntary signal labs use to manage risk. Which outcome follows will decide whether this shifts AI governance or only how individual companies behave.
