Anthropic built an AI that can uncover large swathes of software flaws — and has chosen not to release it publicly. The company says its new model, Mythos, can surface large numbers of vulnerabilities across major operating systems and web browsers, and is sharing access only with a shortlist of large firms — from cloud providers to banks — so they can patch exposed weaknesses before adversaries can exploit them.

What Anthropic did and why

Anthropic announced that Mythos can surface large numbers of software vulnerabilities and that the risks of a public release are too high. Instead of an open launch, the company has restricted distribution to major organisations that operate critical internet infrastructure — including cloud firms, hardware vendors and banks — so they can test and patch exposed weaknesses.

The company has briefed senior government officials and key industry stakeholders and is offering support for official testing and evaluation.

Official concern and industry meetings

US officials treated the development seriously and held calls with executives from leading tech and cybersecurity firms to discuss model security and how to respond if model capabilities scale in favour of attackers.

Separately, senior Treasury and financial regulators met with top bank chiefs to discuss Mythos and AI-related cyber risks to the financial system, signalling government concern about potential systemic exposure in the banking sector.

Security gains and commercial questions

Security teams see a clear upside to restricted access: giving major operators early use lets them find and fix bugs before malicious actors can weaponise the same techniques.

  • Defensive argument: Security executives and some researchers say the speed and scale of attacks will increase if advanced models become widely available to criminals, so defensive preparations need to accelerate.
  • Commercial concern: Others argue the controlled rollout also reinforces a dynamic where frontier models are kept behind enterprise agreements, potentially giving large labs a competitive advantage and making it harder for smaller teams to reproduce capabilities via distillation.

Do smaller models already match Mythos?

Not everyone accepts that Mythos is unique. Some AI cybersecurity startups say they have reproduced aspects of what Anthropic describes using smaller, open-weight models, arguing that success in vulnerability discovery often depends on task framing and tooling rather than on a single dominant model.

Related Articles

Anthropic warned the fallout for economies, public safety and national security could be severe.

This article was created with AI assistance.