Analysis
Pulse reported on September 6 that OpenAI, Google and Anthropic had each shipped a cybersecurity-focused model rated at their own highest internal risk tier within the same two-week window, without publishing the technical detail needed to compare the actual restrictions each company was imposing. In the days since, the labs' actual response has become clearer -- and it isn't more disclosure. It's harder gating.
OpenAI's GPT-6 Astra, unveiled September 3, is described in the company's own materials as the first model to trigger its "Critical" cybersecurity capability threshold under its Preparedness Framework -- OpenAI's internal system for classifying how much a model could plausibly assist a sophisticated attacker. Rather than publish the specific evaluation methodology or the exact restrictions that threshold triggers, OpenAI's answer has been structural: every account on its Daybreak cybersecurity access program, split into Blue (defensive, more widely available) and Red (offensive-capability, tightly restricted) tiers back on August 10, is now required to adopt hardware security keys starting September 1. That's a real operational cost for enterprise customers, not a policy statement -- and it's a tell that OpenAI itself doesn't trust software-only access controls to hold at this capability tier.
What actually changed since Pulse's original coverage
The original story's core complaint -- that none of the three labs had published comparable technical detail -- still stands. What's new is that each lab's practical answer to "how do we prove this is safely restricted" has converged on the same non-answer: tighter gating instead of shared methodology. Google DeepMind's Gemini 3.8 Flash Cyber variant, which followed on September 2, and Anthropic's Mythos 5.1 (Claude's trusted-access twin, shipped September 1 alongside a 75% cut to cache-read pricing) are both access-restricted along similar lines -- vetted defender programs rather than open API access -- without any of the three labs cross-referencing each other's restriction criteria in public.
Reported refusal rates add a second data point: figures circulating put Astra and Anthropic's Mythos access programs at roughly a 91.5% refusal rate on certain cybersecurity-adjacent prompt categories for accounts without the elevated access tier. A refusal rate that high suggests these labs are choosing to hobble the general-access version fairly aggressively rather than build fine-grained technical controls that separate legitimate defensive use from misuse -- which is a defensible safety choice, but a different one than "we've solved the comparison problem," and it pushes the actual risk-management decision behind closed-door vetting criteria that outside researchers and regulators still can't audit.
The practical effect is that the industry has answered a transparency problem with an access-control solution, which solves the misuse-risk question (probably) without solving the comparability question the original story raised. A customer, regulator, or competitor still cannot look at OpenAI's Daybreak Blue/Red split, Google's defenders-only Gemini Cyber variant, and Anthropic's Mythos vetting criteria and determine whether one lab's "Critical" threshold is meaningfully stricter or looser than another's -- they can only observe that all three have made access harder to get, by different and non-comparable mechanisms.
What to watch: whether Massachusetts's proposed catastrophic-risk review requirement, or a similar state-level mechanism, ends up being the thing that forces actual methodology disclosure, since voluntary convergence on tighter gating clearly isn't producing it on its own.
