Anthropic’s ‘Mythos’ AI raises cybersecurity fears, withholds public release over ‘dangers’

Anthropic is sending shockwaves through the tech industry after unveiling a powerful new artificial intelligence (AI) model it says is too dangerous to release to the public.

The system, called Mythos, is being described by the company as a major leap in cybersecurity capable of identifying and potentially exploiting critical software vulnerabilities at unprecedented speed and scale.

According to Anthropic, early testing found the model was able to detect thousands of high-severity flaws across widely used operating systems and web browsers, raising alarms about how such technology could be used if it fell into the wrong hands.

At the center of the concern is the model’s reported ability to autonomously uncover so-called zero-day vulnerabilities, previously unknown flaws that hackers often race to exploit before they can be patched. Security experts warned that widespread access to tools like this could dramatically accelerate the speed and scale of cyberattacks.

In response, Anthropic announced Project Glasswing, an invite-only effort involving select partners, including some competitors, to test the system in controlled environments and strengthen digital defenses.

The development has drawn attention at the highest levels of government and finance. Reports indicate that officials, including Federal Reserve Chairman Jerome Powell, have been briefed on the potential risks associated with rapid advances in AI-driven cybersecurity tools.

Reaction across the tech world has been mixed. Some see the model as a sign that artificial general intelligence, or AGI, may be closer than expected, while others question whether the announcement overstates the system’s capabilities.

Anthropic itself has taken a more cautious stance, saying that while Mythos represents a significant improvement over earlier models, there is no clear evidence it can operate independently at a superhuman level — a key threshold for true AGI.

Still, early testers say the model has already uncovered more vulnerabilities in weeks than they had found over the years, including flaws that could allow attackers to take full control of systems.

Even if this model falls short of a true breakthrough, experts said it offers a glimpse of what’s ahead — a rapidly approaching future where AI tools could empower both cyber defenders and attackers to move faster than ever before.

Follow Luke Duecy on X. Read more of his stories here. Submit news tips here.