“Dangerous” AI models are coming no matter what

late last week, Anthropology took its new form Cloud Fable 5 and Mythos 5 The AI ​​models are offline following US government export control directives that prohibit “any foreign national” from using the services. The company has been in Talks with the White House Since Friday, but it has not yet reached an agreement that would allow it to return the shows.

since Mythos debuted in AprilAnthropic claimed — and warned — that the model has advanced capabilities to not only find vulnerabilities in software to help defenders patch them, but also discover ways to exploit them that could be used by bad actors. Anthropic itself noted this double-edged sword in its launch of Mythos 5 and Claude Fable 5. “Much of the advanced use of AI models is dual use: the same queries that are useful in the hands of cybersecurity professionals and biology researchers can be dangerous if they are available to malicious actors,” the company says. he wrote in a blog post Last week.

With this in mind, the company initially released a version called Mythos Preview to a select consortium as part of a working group known as Project Glasswing. Mythos 5 was also released privately to this group last week, while Claude Fable 5, a Mythos-class model, was released to the general public with a specific ban on its ability to provide answers to questions about biology and cybersecurity.

Then, last weekend, the Trump administration Go to constrain both forms Because she believes that the Fable 5’s guardrails can be disabled to allow full access to the Mythos 5’s capabilities, making her a national security risk.

However, experts say this institutional clash simply delays or hides a harsh truth: Anthropology may be the tip of the spear at this moment, but AI capabilities in general and models from multiple companies and open-weight developers are certain to have capabilities similar to those in Mythos 5 in the near future — if they don’t already have them.

“It is extremely shortsighted to think that no other Anthropic competitor will be able to develop capabilities similar to Mythos or that they haven’t already,” says Tara Wheeler, chief security officer at specialist cybersecurity consultancy TPO Group. “There are other companies following Anthropic’s lead, which may also have capabilities, and are holding them in reserve as they see how Anthropic is being treated in the current regulatory environment.”

Anthropic itself has emphasized this point since launching the Mythos Preview. “The real message is that it’s not about the model or the human,” Logan Graham, leader of the company’s Frontier Red team, told WIRED when the Mythos Preview launched in April. “We need to prepare now for a world where these capabilities are widely available in 6, 12 or 24 months.”

OpenAI, for example, has also made a special release of A model focused on cybersecurity In mid-April it announced an expanded cybersecurity strategy.

The researchers point out that even before this next generation of models, existing AI offerings could be used for advanced searches for vulnerabilities and exploit them with an improved tool. A large group of cybersecurity leaders confirmed this to management in a meeting Open letter on Sunday, arguing that White House guidance on export controls was misguided.

“It’s not one model; it’s the general trend of technology,” says Bruce Schneier, a researcher at Harvard University and the University of Toronto who taught at Harvard. analysis the situation. “Smaller, cheaper, open source models can, sometimes alone and sometimes in concert with each other, match Mythos/Fable’s performance with a more complex claim. We should expect other models to match Mythos/Fable’s creativity and tenacity within months – a little longer for open source models.”

What the White House and governments around the world need to focus on, experts say, is democratically developing broader, more transparent plans for how to handle advances in AI capabilities in cybersecurity and other sensitive areas when they inevitably occur.

“The policy question is not whether the technology is risky,” says Chris Wysopal, co-founder of cloud security company Veracode. “The question is whether a particular restriction meaningfully reduces that risk or whether it fundamentally slows down people trying to make systems more secure.”