OpenAI’s GPT-5.5 Cyber Goes Restricted After Anthropic’s Mythos Kerfuffle

Remember when Anthropic took flak for locking down Mythos, their specialized cybersecurity model, to only a select group of trusted defenders? Well, history just repeated itself. OpenAI announced today that GPT-5.5 Cyber, their new cybersecurity testing tool, will roll out exclusively to “critical cyber defenders” first.

This isn’t entirely surprising. Both companies are walking the same tightrope: they want to empower the good guys without handing ammunition to the bad ones. But the timing feels off. Anthropic caught heat for being too restrictive with Mythos, and now OpenAI is essentially doing the same thing—just with a slightly different label.

Let’s be real: “critical cyber defenders” is vague enough to make any security researcher raise an eyebrow. Who decides who’s critical? How do you verify that? OpenAI hasn’t detailed the vetting process, which leaves room for plenty of skepticism. I’ve seen this play out before with other restricted-access tools, and the result is often a black market for access or frustrated researchers who end up building their own alternatives.

The model itself is impressive on paper. GPT-5.5 Cyber is designed to simulate attacks, find vulnerabilities, and even suggest patches—all tailored for cybersecurity pros. But if only a handful of people can actually use it, what’s the point? The whole value of an AI tool like this is scale. Limiting it to a few hundred defenders might make for good PR, but it doesn’t move the needle on the global threat landscape.

What’s more interesting is the contrast with OpenAI’s earlier stance. They’ve been vocal about democratizing AI access, yet here they are, locking down a tool that could genuinely help small teams and independent researchers. It feels like a backtrack, even if the reasoning is sound. I get the safety concerns—nobody wants a rogue state using GPT-5.5 Cyber to crack critical infrastructure—but the execution feels half-baked.

I’d rather see a tiered approach: give full access to vetted defenders, but offer a sandboxed version with limited capabilities to everyone else. That way, you still control the dangerous stuff while letting the community experiment and learn. Right now, this feels like a gatekeeping move that benefits no one except the PR department.

Anthropic’s Mythos restrictions sparked a debate that’s still smoldering. OpenAI just threw more fuel on the fire. The real question is whether either company will listen to the backlash, or if we’re stuck with a future where cutting-edge AI tools are only for the “chosen few.”