Anthropic is not releasing its most powerful AI. The model, codenamed Mythos, can find 27-year-old bugs in OpenBSD and map attack paths through core financial and energy systems. Dario Amodei’s team isn’t selling it. They’re rationing it to just 11 systemically important firms - Apple, JP Morgan, Microsoft - and the UK government.

The containment strategy failed. According to Saagar Enjeti on Breaking Points, a hacker collective on Discord now has access to the model. That breach turns Anthropic’s safety-first posture into a cautionary tale: one server crack can unleash AI-driven cyberattacks at industrial scale.

"The Bank of England has already warned that Mythos could crack the entire cyber risk world open."

Saagar Enjeti, Breaking Points with Krystal and Saagar

Alex Hearn on The Intelligence argues the gatekeeping serves more than safety. It masks a compute crunch and blocks Chinese rivals from cloning the model’s outputs. But the deeper shift is structural: AI has moved from writing code to finding zero-days at superhuman speed. If Mythos can dismantle OpenBSD, it can dismantle JPMorgan’s firewall.

Governments are unprepared. Canada’s finance minister likened the threat to a blockade of the Strait of Hormuz - same strategic impact, no military action required. Yet no agency regulates AI models that can collapse critical infrastructure. Medical drugs face years of trials. AI tools face a startup CEO’s judgment.

"We are betting the stability of the global financial system on the server security of a single company."

Krystal Ball, Breaking Points with Krystal and Saagar

Cat Wu, Head of Product for Claude Code, says Anthropic ships features in 24-hour sprints. But when the feature is a model that can automate systemic attacks, speed becomes a liability. The leak proves that even the most safety-obsessed lab can’t secure what it chooses to build. The question isn’t whether Mythos should be released. It’s whether anyone should have been allowed to build it.