Anthropic is raising concerns about the new AI model, ‘Claude Mythos’, calling attention to its potential dangers. Executives have warned that if this model is made widely available, it could lead to severe hacking incidents and even terrorism.
In a rather unsettling assessment, Anthropic noted that, in the wrong hands, Mythos could target crucial infrastructure like power grids and hospitals. The company claimed that the model has already identified thousands of serious vulnerabilities across major operating systems and web browsers.
Instead of launching it publicly, Anthropic, under CEO Dario Amodei, plans to implement “Project Glasswing.” This initiative will provide access to a limited group of around 40 companies—such as Amazon, Google, Apple, Nvidia, CrowdStrike, and JPMorgan Chase—allowing them to use Mythos for identifying and addressing security issues.
Roman Yampolsky, an AI safety researcher at the University of Louisville, commented that this controlled rollout might be the best strategy for Anthropic, as it would empower those who can patch vulnerabilities while keeping it away from potential hackers.
“I suppose there will be some leaks,” he remarked, adding, “Any restrictions are better than unrestricted access. Ideally, this model wouldn’t be developed at all.” Yet, it seems development isn’t stopping anytime soon.
Yampolsky further expressed worries about the implications of such models, suggesting they could evolve to create advanced hacking tools and even unprecedented types of weapons.
A particular instance from Anthropic’s tests demonstrated Mythos successfully breaching a “sandbox” designed to limit internet access. This discovery came about quite unexpectedly—during a lunch break in a park. In another example, Mythos uncovered a flaw in the OpenBSD operating system that had remained undetected for nearly three decades.
Despite acknowledging the risks, Anthropic believes that Project Glasswing will bolster U.S. defense capabilities, especially in light of increased threats from adversaries like Iran, China, and Russia.
Anthropic stated its focus is on companies that hold significant portions of the global cybersecurity landscape. “These organizations maintain operating systems, browsers, and financial infrastructures that billions rely on,” an official noted. “When vulnerabilities are patched, it safeguards everyone using that software, often protecting hundreds of millions.”
Discussions are ongoing with U.S. government representatives regarding how Mythos can enhance national cyber capabilities both defensively and offensively.
Elia Zaitsev, chief technology officer at CrowdStrike, remarked that the Claude Mythos Preview reveals what defenders can achieve at a larger scale, yet adversaries will seek to exploit similar capabilities.
However, some critics are doubtful that Anthropic’s public approach aligns with its stated commitment to safety. Perry Metzger, president of the AI policy group Alliance for the Future, pointed out the intense buzz surrounding Mythos that has stemmed from the company’s warnings.
“Either you pay for access to Glasswing or you simply don’t participate—it’s as if they’re saying they alone decide who should get access,” Metzger stated, adding a hint of sarcasm. “It’s frustrating.”
According to reports, some of Anthropic’s critics, including former Trump AI advisor David Sachs and White House officials, view the company’s safety alerts as an attempt at “regulatory cheating,” a term that implies manipulating rules to favor the company while disadvantaging competitors.
Nathan Riemer, executive director of advocacy group Build American AI, commented on Amodei’s approach, suggesting he sees himself as the sole authority in determining the technology’s potential and who can access it. “Who needs a public square, after all?”
Another tech industry professional, who chose to remain anonymous, recalled a similar situation with OpenAI, which in 2019 had announced that an early version of its chatbot, GPT-2, was too risky to release. Notably, Amodei and Jack Clark were both involved with OpenAI back then.
There’s speculation that the limited rollout might be more about Anthropic facing challenges in meeting computing demands rather than genuine safety concerns. “They might be trying to divert attention from their capacity issues,” the source hinted.
