AI could learn to break any security within months

New neural network models may become capable of "breaking any security system" within the coming months, Report informs via The Guardian, citing a warning from the Five Eyes intelligence alliance.

The alliance includes the United States, the United Kingdom, Canada, Australia, and New Zealand. Western intelligence agencies warned that new artificial intelligence models pose an urgent cybersecurity threat and should be treated as a serious risk for businesses.

According to the information, AI is learning not only to strengthen corporate defenses but also to scale cyber threats, accelerate attacks, and make them increasingly sophisticated.

The newspaper notes that members of the alliance rarely issue public warnings. The statement came after the US administration reportedly instructed Anthropic to restrict foreign users' access to its most powerful AI models, Fable 5 and Mythos 5.

Anthropic's advanced AI model, Mythos, was able during a controlled security evaluation to "break into almost everything" within the secret systems of the US National Security Agency (NSA)-one of the country's highest-level intelligence organizations-in just a few hours.

However, the model merely analyzed the codebase to complete assigned tasks and discovered several previously unknown minor vulnerabilities. No actual offensive hacking activity or real-world intrusion took place.