Anthropic CEO Dario Amodei warns that AI safety is no longer theoretical. A new tier of models, termed Mythos-class, can now identify and exploit thousands of software vulnerabilities within hours. This capability fundamentally alters the cybersecurity landscape by weaponizing flaws at machine speed.
Amodei estimates Chinese AI models currently trail US advancements by six to twelve months. However, he cautions that China’s aggressive open-source ecosystem is closing this gap rapidly. Once powerful model weights are released publicly, safety guardrails become optional and US export controls lose effectiveness. This creates a direct threat vector for critical infrastructure and national security.
In response, the US Commerce Department imposed export controls in June 2026 targeting Anthropic’s advanced Fable 5 framework. These measures aim to prevent unauthorized access to dangerous AI tools. Anthropic has simultaneously restricted Mythos capabilities through a controlled internal initiative known as Project Glasswing.
For investors, this geopolitical friction introduces significant regulatory risk. Export controls on advanced models may limit revenue opportunities for firms with international exposure. If Chinese labs achieve Mythos-class capabilities by early 2027, the global cybersecurity and market access environment will shift dramatically.