‘Dangerous’ AI Models Are Coming No Matter What

Late last week, Anthropic took the new Claude Fable 5 and Mythos 5 AI models offline to comply with the United States government’s law prohibiting “any foreigner” from using the services. The company has entered he talks to the White House as of Friday but have yet to reach an agreement that would allow them to return the donation.

Since The stories started in AprilAnthropic said—and warned—that the model has a high level of ability not only to find software vulnerabilities to help security guards patch, but also to find exploits that can be used by malicious actors. Anthropic itself saw this double-edged sword when launching Mythos 5 and Claude Fable 5. “The main use of AI models is two-fold: the same questions that are beneficial in the hands of cybersecurity experts and biological researchers can be dangerous if they exist for criminals,” the company. he wrote in a blog post last week.

With this in mind, the company initially released a version called Mythos Preview to a select group as part of a working group called Project Glasswing. Mythos 5 was also released privately to the team last week, when Claude Fable 5, which is a Mythos-grade version, was released to the general public with specific blocks on its ability to answer questions about biology and cybersecurity.

Then, at the end of last week, the Trump administration moved to ban both types because they believe that Fable 5’s guardrails can be disabled to allow full access to Mythos 5’s power, it is said to be causing a threat to national security.

Experts say, however, that this institutional debate is only delaying or hiding a hard truth: Anthropic may be the tip of the spear at this time, but the power of AI in general and models from several companies and developers of open weight will have the same power as Mythos 5 in the near future – if not.

“It’s very difficult to imagine that any other Anthropic competitor will have the same capabilities as Mythos or even if they haven’t already,” said Tarah Wheeler, chief security officer at the private security research firm TPO Group. “There are other companies that are on Anthropic’s heels that probably have potential, too, and they’re keeping them as they see how Anthropic is doing in the areas they’re managing.”

Anthropic itself has emphasized this point since the launch of the Mythos Preview. “The real message is that this is not about color or Anthropic,” Logan Graham, the company’s red team director, told WIRED when the Mythos Preview was launched in April. “We have to prepare for a world where this technology is available in 6, 12, 24 months.”

OpenAI, for example, also privately released a example of cybersecurity in April and announced an expanded cybersecurity strategy.

Researchers note that even before this next generation, existing AI offerings can be used for vulnerability detection and exploitation in development with refined tools. A large group of cybersecurity leaders emphasized this to managers in the an open letter On Sunday, he argued that the White House’s directive to drive abroad was wrong.

Bruce Schneier, a researcher at Harvard University and the University of Toronto, said: “It’s not just one example; analysis the situation. Small, low-cost, open-ended models, sometimes independently and sometimes in combination, can match Mythos/Fable systems with high-level concepts.

What the White House and governments around the world need to focus on, experts say, is creating broader and more transparent democratic policies on how to deal with advances in AI in cyber security and other critical areas where it happens.

“The policy question is not whether technology is at risk,” says Chris Wysopal, co-founder of cloud security firm Veracode. “The question is whether a particular ban reduces that risk or significantly reduces the number of people who want to keep machines safe.”