Krux
March 28, 2026
Anthropic's Leaked Model Is Too Good at Hacking
Published: March 28, 2026 at 12:38 AM
Updated: March 28, 2026 at 12:38 AM
100-word summary
Anthropic confirmed it's testing Claude Mythos after a data leak revealed the model's existence. It's their most capable AI yet, sitting in a new tier above Opus called Capybara. Mythos scored dramatically higher on coding, reasoning, and cybersecurity benchmarks. Here's the catch: it's so good at finding security vulnerabilities that Anthropic is keeping it locked down to a handful of early customers. The model is expensive to run and reportedly capable of actual cyber exploits, not just spotting them. What Anthropic built is powerful enough that releasing it widely feels riskier than keeping it secret.
What happened
Anthropic confirmed it's testing Claude Mythos after a data leak revealed the model's existence. It's their most capable AI yet, sitting in a new tier above Opus called Capybara. Mythos scored dramatically higher on coding, reasoning, and cybersecurity benchmarks. Here's the catch: it's so good at finding security vulnerabilities that Anthropic is keeping it locked down to a handful of early customers. The model is expensive to run and reportedly capable of actual cyber exploits, not just spotting them.
Why it matters
What Anthropic built is powerful enough that releasing it widely feels riskier than keeping it secret.