Anthropic CEO Dario Amodei: In 6-12 months, China's open-source AI models will be able to catch up to Mythos

Anthropic CEO Dario Amodei publicly revealed in a “Lunch with the FT” interview that open-source models and Chinese developers could replicate Mythos’s cybersecurity capabilities in as fast as 6 months, and at most 12 months.
(Mythos is Anthropic’s most advanced, unreleased model.)
(Background: Anthropic Mythos is so powerful that it triggered an emergency meeting: Bencet and Powell summoned Citibank, Goldman Sachs, Bank of America, and JPMorgan Chase, with the five major banks focusing on financial risks.)
(Additional context: Short seller Michael Burry announced that Palantir is just a low-margin SaaS outsourcing company! Anthropic is eating it up.)

In 6 to 12 months, Dario Amodei gave a very exaggerated countdown in the Financial Times.

He said that open-source models and Chinese developers could, within this timeframe, replicate Mythos’s cybersecurity capabilities, which is a kind of war signal.

What can Mythos do? Anthropic claims that it previously identified thousands of zero-day vulnerabilities across “every major operating system and browser,” including one that had been hidden in OpenBSD for a full 27 years, which no one had discovered before.

In the past two weeks, the White House has promoted federal agencies to adopt Mythos, with Chief of Staff Susie Wiles directly meeting with Amodei on April 17; Citibank, Goldman Sachs, Bank of America, Morgan Stanley, and JPMorgan Chase were all urgently summoned to assess financial infrastructure risks.

Amodei’s strategic logic is that if the spread of model capabilities cannot be prevented, the sequence can be arranged. Before releasing to the public, defenses should be pushed to the limit.

Dario Amodei’s warning

Anthropic’s internal red team leader Logan Graham previously stated that other companies could replicate Mythos-level capabilities within a few months, while open-source efforts in China would take 1 to 1.5 years.

Dario Amodei estimates that the 6-12 month timeframe is somewhat optimistic. He does not believe high-performance model capabilities can be permanently locked down; what matters is how many existing system vulnerabilities can be patched before this time arrives.

Dario Amodei currently treats Mythos as a nuclear weapon—everyone has the ability to create it, but strict control over who can access it is necessary.