Krux
April 9, 2026
Anthropic Built an AI Hacker Too Dangerous to Release
Published: April 9, 2026 at 12:34 AM
Updated: April 9, 2026 at 12:34 AM
100-word summary
Anthropic's new model Mythos Preview can autonomously find thousands of serious security flaws in major operating systems and browsers, then write working exploits for them. That's powerful enough that the company won't release it publicly. Instead, only 50-plus organizations (Apple, Microsoft, AWS, JPMorgan) get access to use it for defensive security testing. The coalition gets $100 million in usage credits to hunt vulnerabilities in their own systems before attackers do. Anthropic will donate $4 million to open-source security projects and publish findings within 90 days. It's a rare admission: we built something genuinely dangerous, so we're treating cybersecurity like a shared infrastructure problem rather than a product launch.
What happened
Anthropic's new model Mythos Preview can autonomously find thousands of serious security flaws in major operating systems and browsers, then write working exploits for them. That's powerful enough that the company won't release it publicly. Instead, only 50-plus organizations (Apple, Microsoft, AWS, JPMorgan) get access to use it for defensive security testing. The coalition gets $100 million in usage credits to hunt vulnerabilities in their own systems before attackers do. Anthropic will donate $4 million to open-source security projects and publish findings within 90 days.
Why it matters
It's a rare admission: we built something genuinely dangerous, so we're treating cybersecurity like a shared infrastructure problem rather than a product launch.