Story perspectives

Anthropic Limits Claude Mythos Over Autonomous Zero-Day Exploits

4/24/2026

Summary

Anthropic Limits Mythos

Anthropic limited Claude Mythos to select partners because it autonomously exploits zero-day vulnerabilities.
The model finds software flaws and has an 83 percent first-try exploit success rate.
Project Glasswing preview lists Apple, Microsoft, Google, Amazon Web Services, Nvidia, JPMorgan Chase and the NSA.
Anthropic pledged US$100 million in usage credits and US$4 million in open-source grants.
The World Economic Forum warned AI-driven attacks could outpace patch cycles.

Summary

Anthropic Investigates Breach

Anthropic said a Discord server accessed Claude Mythos by guessing its location with data from a Mercor breach and a contractor’s credentials.
The intrusion entered through a third-party vendor environment where a member held evaluation permissions for Anthropic models.
Anthropic is investigating and has found no evidence the breach spread beyond the vendor’s environment.
Security researchers Lukasz Olejnik and Pia Hüsch called the incident an “entirely imaginable” supply-chain failure.

History

Historical Context

The 2013 Target data breach, which compromised 40 million credit card numbers, underscored vulnerabilities in supply chain security, a concern heightened post-9/11. As organizations increasingly depend on third-party vendors, the need for robust cybersecurity measures has grown. Recent incidents, like Anthropic's breach, illustrate ongoing challenges in securing AI technologies, emphasizing the critical importance of safeguarding data integrity in modern supply chains.

1 of 3

Anthropic Limits Mythos

Anthropic limited Claude Mythos to select partners because it autonomously exploits zero-day vulnerabilities.
The model finds software flaws and has an 83 percent first-try exploit success rate.
Project Glasswing preview lists Apple, Microsoft, Google, Amazon Web Services, Nvidia, JPMorgan Chase and the NSA.
Anthropic pledged US$100 million in usage credits and US$4 million in open-source grants.
The World Economic Forum warned AI-driven attacks could outpace patch cycles.

1 / 3