Two labs built two different answers to the access question. This week, both got the same veto.
Focus 1 has spent five posts on a real divide: Anthropic grants access to a short list of named partners through Project Glasswing, while OpenAI has built a tiered, verification-based structure that scales past what a partner list can handle. The original draft of this post treated that divide as the whole story, and argued OpenAI's model was the more stable one because it hadn't had a public failure the way Anthropic's had. That argument doesn't survive this week. On July 9, OpenAI brought GPT-5.6 Sol, Terra, and Luna to general availability, weeks later than planned, after holding the release for a government safety review tied to the models' ability to find code vulnerabilities. The formal mechanism behind that review is a June 2 executive order letting developers give the government up to thirty days of access to a covered model before releasing it to trusted partners, under a framework agencies were still finalizing. OpenAI's trusted-partner release happened on June 26, which is where that window closes on the order's own terms. The thirteen days between that release and the July 9 public rollout sit outside what the order formally governs. Something still shaped that gap. It just wasn't the voluntary review provision itself. That is not the tiered access system this series already covered. It is a second gate, sitting above it, and a less formal one than it first looks.
The Anthropic parallel is closer than a coincidence. On June 12, the Commerce Department imposed export controls on Claude Fable 5 after Amazon researchers reported a way to prompt the model into identifying software vulnerabilities and, in one case, generating exploit code. Anthropic pushed back on the severity of the finding, noting that several less capable models could do the same thing, but the restriction still landed: foreign nationals, including some of Anthropic's own employees, lost access while the government reviewed the case. It took roughly two and a half weeks of negotiation before Fable 5 returned to general availability on July 1, with new safety classifiers in place. Mythos 5, the more capable sibling, came back only for a vetted subset of Project Glasswing partners, exactly where it started.
Line the two episodes up and the pattern is not about access philosophy anymore. It is about who holds the override. Anthropic's partner model and OpenAI's tiered model fail differently when the vendor is the only party deciding, the way F1-P1 through F1-P5 have argued. But both models just demonstrated they can be overridden by a party that isn't the vendor and isn't the customer at all. The government didn't ask Anthropic to renegotiate a partner agreement or ask OpenAI to adjust an authentication tier. It reviewed each model on its own timeline, for reasons neither company fully controlled.
That is the detail enterprise security teams need to sit with. A revocation risk you can at least see coming, because it is written into a contract or a tier policy, is a planning problem. A government review with no published criteria and no fixed timeline is not. Anthropic has since committed to four things meant to make this less arbitrary going forward: expanded pre-release access for government evaluators, threat intelligence sharing on how its tools get abused, participation in a government-run vulnerability clearinghouse, and work with Amazon, Google, and Microsoft on a shared framework for rating jailbreak severity. Whether OpenAI adopts anything comparable after this week's delay is worth watching. If it does, that starts to look less like two labs handling a crisis separately and more like the outline of a standard neither company chose on its own.
For now, the operating reality is blunt. A security team evaluating which frontier lab to build on cannot treat the vendor's access philosophy as the only variable that matters. Government review has now interrupted both major access models this year, and neither Anthropic's named-partner structure nor OpenAI's tiered-verification structure prevented it. The question this series opened with was who owns the access decision. The honest answer, after this week, is that neither lab fully does.
Closing Signal: This is the sixth post in Weaponized Access — Who Owns the Decision? The series asked who controls the access decision. This week, both labs got the same answer from the same source. The full account of how the government's review reached each company, including the sourcing behind it, is in the deep dive.
