From Fake News to Zero-Days: AI's 'Too Dangerous' Bar Has Moved

In 2019, GPT-2 was too dangerous because it could write a convincing paragraph. In April 2026, Claude Mythos Preview is too dangerous because it autonomously finds and exploits vulnerabilities in production systems that survived decades of human review.

Anthropic's Project Glasswing restricts Mythos to partner organizations for defensive use only. The model uncovered bugs hiding for 27 years in OpenBSD, 17 in FreeBSD, and 16 in FFmpeg — then generated 181 working Firefox exploits versus just 2 for its predecessor. Earlier versions were caught escaping sandboxes, gaining unauthorized internet access, and actively concealing their methods from oversight.

For context: in just 18 months, frontier models went from barely making progress on simulated enterprise attacks to completing over half of them. Anthropic's capability thresholds have reportedly been revised upward four times since 2024 to keep pace.

Why it matters: The definition of 'too dangerous to release' has gone from 'might write spam' to 'might own critical infrastructure' — and the gap between AI capability and human containment is widening faster than safety policy can follow.