Anthropic just told the US government that their unreleased AI model could fuel large-scale cyberattacks in 2026. Let me be clear: this is either the most responsible thing a tech company has done in years, or the most elaborate marketing stunt I’ve ever seen.
The company announced this week that Claude Mythos—their latest AI model—is too good at hacking to release to the public. According to Anthropic, this thing represents a “step change” in performance, which in AI-speak means it’s not just incrementally better. It’s different.
What We Actually Know
The facts are thin, but what we have is concerning. Anthropic has decided to keep Mythos under strict control due to its advanced hacking capabilities. They’re not releasing it. Period. The company is testing it internally, but you and I won’t be getting our hands on it anytime soon.
This came to light after an accidental data leak revealed the model’s existence. Nothing says “we have this under control” quite like finding out about a dangerous AI through a leak, right?
The Containment Problem Nobody Wants to Talk About
Now, I’ve tested dozens of AI models for this site. I’ve seen them fail in hilarious ways. I’ve seen them succeed in terrifying ways. But I’ve never seen a company flat-out refuse to release a model because it’s too capable.
This raises an uncomfortable question: if Mythos is too dangerous to release, what does that say about the models we’re already using? Claude 3.5 Sonnet is publicly available right now. Is it safe because it’s actually safe, or just because it’s not quite dangerous enough to worry about?
The cyberattack angle is particularly interesting. We’re not talking about a model that might occasionally help someone write malicious code. Anthropic is warning government officials about potential large-scale attacks in 2026. That’s specific. That’s scary.
Why This Matters for Regular Users
You might think this doesn’t affect you because you’ll never use Mythos. Wrong. This decision sets a precedent for how AI companies handle capability thresholds. If Anthropic can identify a model as too dangerous and actually hold it back, that’s new territory.
Most AI labs have been in an arms race to release the most powerful model possible. OpenAI releases GPT-4, Google counters with Gemini, Anthropic drops Claude 3. Everyone’s trying to one-up each other. This is the first time I’ve seen a major player pump the brakes.
But there’s a cynical read here too. Anthropic gets to claim they built something so powerful they can’t release it, which is fantastic PR. They look responsible and capable at the same time. Meanwhile, we have no way to verify any of these claims because, conveniently, we can’t test the model.
The Testing Question
Anthropic says they’re testing Mythos internally. Great. Who’s doing that testing? What are the protocols? What happens if it “breaks containment” during testing—a phrase that sounds like it came from a sci-fi movie but is apparently our reality now?
The company has decided to prevent potential misuse by keeping the model locked down. That’s the right call if Mythos is genuinely as dangerous as they claim. But it also means we’re taking their word for it. There’s no independent verification. No red team from outside the company. Just trust us, they say.
What Happens Next
If Anthropic is serious about this, other AI labs will face pressure to adopt similar standards. That could be good. Or it could mean we end up with a two-tier system: safe models for the public and dangerous models for governments and corporations.
I’ve spent years reviewing AI tools, and I’ve learned one thing: companies rarely hold back their best work unless they absolutely have to. If Anthropic is keeping Mythos locked up, I believe them that it’s dangerous. What I don’t know is whether they’re doing enough to keep it contained, or whether this whole situation is a preview of much bigger problems ahead.
For now, Claude Mythos remains in Anthropic’s vault. Whether it stays there is anyone’s guess.
🕒 Published: