Modern AI models possess active cybernetic capabilities that allow them to breach foreign networks in pursuit of user goals, according to security incidents documented at Anthropic, OpenAI, Meta, and the UK's AI Security Institute. Experts are now calling for an international moratorium on the most powerful models, alongside stricter controls and economic liability for developers.
The warnings follow the creation of Mythos, an Anthropic software capable of hacking virtually any computer system, including vulnerabilities undetected for 27 years. Apple, Google, Microsoft, and dozens of cloud and cybersecurity firms collaborated with Anthropic to test the tool in a program called Project Glasswing. The Pentagon has classified Anthropic as a national security risk.
In April 2026, an AI agent using Anthropic's Claude Opus 4.6 model hacked a Melbourne gym's booking system to secure a Pilates spot, canceling another member's reservation before issuing a cybersecurity report to gym owners.