OpenAI’s GPT-5.5 Cyber: Another Restricted Release, Another Double Standard

OpenAI just announced it’s rolling out GPT-5.5 Cyber, their cybersecurity testing tool, but only to “critical cyber defenders” at first. Which sounds noble until you remember how loudly they complained when Anthropic did the same thing with Mythos.

Remember that? A few months back, OpenAI was on a high horse about Anthropic limiting Mythos access to select researchers and government agencies. Called it anti-competitive, said it stifled innovation, the whole nine yards. Now here we are, with OpenAI pulling the exact same move on Cyber.

The justification is predictable: safety concerns, preventing misuse, ensuring the tool doesn’t fall into the wrong hands. And look, I get it—cybersecurity tools are double-edged swords. Give a powerful AI-driven penetration tester to the wrong people and you’re basically handing out skeleton keys to every system it can reach.

But here’s the thing: when Anthropic did this, OpenAI’s argument was that broad access would democratize security research and let smaller teams compete with well-funded attackers. That argument hasn’t changed. The threat landscape hasn’t changed. What changed is whose ox is being gored.

What counts as “critical cyber defenders” anyway? OpenAI hasn’t defined it clearly. Is it just government agencies and Fortune 500 security teams? What about the small consultancy that protects regional hospitals? Or the open-source tool maintainer who patches vulnerabilities for free? They’re probably out of luck.

This is higher than I expected in terms of hypocrisy, even for this industry. The pattern is getting old: criticize competitors for restrictive access, then adopt the same restrictions once you have a competing product. It’s not about safety—it’s about controlling the narrative and the market.

I’m not saying GPT-5.5 Cyber should be handed out like candy. But if OpenAI truly believed broad access was better for security, they’d walk the walk. Instead, they’re giving us the same gatekeeping they condemned, just with a different logo on the announcement.

The real question is whether this tool even works as advertised. We’ll find out eventually, but for now, it’s locked behind a vague definition of who qualifies. If you’re not already plugged into the right circles, you’re probably not getting in.

Personally, I’d rather see a tiered approach: full access for verified researchers, limited access for everyone else, with clear criteria and an appeals process. But that would require admitting the previous criticism was hollow, and we all know how that goes.