I read the retelling first: OpenAI was supposedly shipping a cyber model with its safeguards stripped, all but built for writing exploits. Then I opened the original. No such sentence exists there.
Fact
The facts: on August 10 OpenAI announced an expansion of the Daybreak Cyber Partner program. Frontier cyber models go to security partners: Accenture, IBM, Capgemini, Cognizant, EY, KPMG, PwC, NCC Group, SpecterOps. The officially stated goal: finding vulnerabilities, judging which ones are a real threat, and fixing them faster, inside the systems defenders already use. The wording on access: more access, with the safeguards to match. There is no CVE: this is a model access program, not a software flaw.
Why a partner program instead of an API for everyone? Because this class of models is dual-use by birth. The same capability that finds the hole for the defender finds it for the attacker. A vetted door is the price of that knowledge.
And the gap between the aggregator headline and the official text is the gap between news and bait. One sells a click with fear, the other describes a partner program with audits. They are not the same story.
The partner list is also a list of the accountable. When access has names on it, there is someone to ask.
The original is always two clicks away. Open it.
Official primary sources
OpenAI - Putting frontier cyber models in more trusted handsThe visual is generated code-art, no third-party images.