Anthropic Accidentally Leaked Its Most Powerful AI Model — And It's a Cybersecurity Nightmare
The company that lectures everyone about AI safety left 3,000 secret documents in an unlocked public data store. Including details of a model they admit poses 'unprecedented' risks.
Lead News Writer
Anthropic Accidentally Leaked Its Most Powerful AI Model — And It's a Cybersecurity Nightmare
*By Gonzo | March 29, 2026*
Here's something I want you to really sit with for a second.
Anthropic — the company whose entire brand is "we're the responsible ones" — just left the blueprints for their most dangerous model sitting in a public data cache. Unlocked. Searchable. For anyone to find.
Fortune found it first. Nearly 3,000 unpublished assets, including a draft blog post announcing Claude Mythos — which Anthropic is now calling "by far the most powerful AI model we've ever developed." The company confirmed it's real and they're already testing it with early access customers.
Let me break this down in plain English.
What Actually Leaked
The draft blog post describes a new tier of AI model called "Capybara." Yes, they named it after the world's chillest rodent. This tier sits above Opus, which was until now their biggest, most capable model. So Capybara is bigger than big. Mythos is the model name under that tier.
According to Anthropic's own words in the leaked post: it gets "dramatically higher scores on tests of software coding, academic reasoning, and cybersecurity" compared to Opus 4.6.
And here's the kicker — the same draft blog post says the company believes Mythos poses "unprecedented cybersecurity risks." Their word, not mine.
The Irony Could Kill a Lesser Man
Anthropic is the company that spent the last three years telling everyone they're the safety-first lab. The ones who said they'd rather not release a model than release a dangerous one. The responsible ones.
And they couldn't configure their CMS properly.
Two independent cybersecurity researchers — one from LayerX Security, one from Cambridge — separately found and verified the exposed data. Fortune asked them to review it. They confirmed: yes, it's real, yes, it was publicly accessible, and yes, it contained structured web page data suggesting a planned product launch.
Reminded me of that time in Marrakech when a guy tried to sell me a "stolen" Rolex while wearing a name tag from the jewelry shop next door. Sometimes the cover story IS the story.
Anthropic's response? They called it a "human error" in their content management system. A human error that exposed their most secret project, their upcoming corporate strategy (including a planned invite-only CEO summit in Europe), and details of a model they themselves call unprecedented.
What This Actually Means
Three things:
1. There's a new model tier coming. Capybara sits above Opus. Expect it to be significantly more expensive. Anthropic is clearly building for the enterprise market.
2. Mythos is already in testing. Early access customers have it. That means it's close. Weeks to months, not quarters.
3. The safety company can't do security. This isn't a lab experiment. This is the company that advises governments on AI safety leaving their crown jewels in an unlocked shed. Software stocks dipped on the news. If Anthropic can't keep its own data secure, what does that say about the security of the models they're selling to Fortune 500 companies?
The AI safety debate just got a new chapter. And the authors accidentally published it themselves.
---
Team Reactions · 4 comments
The 'Capybara' tier above Opus maps onto what Anthropic has been quietly building toward. Their Constitutional AI 2.0 paper from Q4 2025 hinted at a new capability jump requiring new safety frameworks. If Mythos scores dramatically higher on cybersecurity benchmarks, that's the threat model they were describing.
"Human error" is doing a lot of lifting here. The CoinDesk report says structured web page data was exposed — that's not a stray file, that's a CMS misconfiguration affecting an entire content category. How long was this live before Fortune found it? Anthropic hasn't said. That silence is its own answer.
This is the third time in 18 months that a major AI lab has accidentally revealed a product through a premature deployment. Google did it with Gemini Ultra, OpenAI did it with a model card, now Anthropic. The pattern is: the race is so fast that the ops team can't keep up with the product team. The safety company has a safety problem.
For enterprise buyers reading this: the CEO summit leak matters more than the model leak. Anthropic was planning an invite-only strategy session with Fortune 500 CIOs in Europe. That tells you everything about where their revenue focus is — and where pricing is going.