Anthropic has warned that the system is so advanced that it is too dangerous to make accessible to the public.
Mythos, Anthropic's artificial intelligence, is transforming the cybersecurity landscape and worrying governments and corporations alike.
Anthropic, a leading company in the sector, has introduced a new release that has set off alarm bells. This is Mythos, an AI model announced last month capable of performing tasks that were previously unimaginable. It is bittersweet news, as the company itself has warned that the system is so advanced that it is too dangerous to make accessible to the public.
The model possesses the ability to act in "hacker mode" fully autonomously, detecting and exploiting security loopholes more effectively than a human. Consequently, its creation poses a global threat; with Mythos, any individual could potentially hack any computer system, including nuclear and power plants.
In fact, according to Bloomberg News, citing documentation and an individual close to the matter, a small group of unauthorized users has managed to gain access to the new Mythos AI model.
As noted in a report, this group belongs to a private online forum and managed to access the system the same day Anthropic announced its plan to launch Mythos only to a limited number of companies for testing purposes. Since then, according to the document, this group has used the model regularly, though not for cybersecurity purposes. "We are investigating a report alleging unauthorized access to Claude Mythos Preview via one of our third-party provider environments," an Anthropic spokesperson stated.
In countries like China and Russia, the arrival of this new AI model has generated unease. According to The New York Times, a Kremlin-aligned Russian media outlet described Mythos as something "worse than an atomic bomb."
However, Anthropic's model isn't the only problem. The company has pointed out that within six to eighteen months, other AI organizations will likely build models similar to Mythos. Furthermore, they added that they "see no reason to think that Mythos Preview is the point where the cybersecurity capabilities of language models reach their limit."
An AI "Too Dangerous"
In just a few minutes, Mythos Preview has the capacity to identify weaknesses and hidden flaws in the IT environments and software of all types of companies and organizations. Because of this, global concern has taken root, largely due to the risk it poses to governments on a global scale. It has also triggered debates among financial institutions, lawmakers, and regulators regarding the threat this AI poses to digital services.
"The risk is not this specific model. The risk is what it represents: that the ability to find and exploit expert-level vulnerabilities no longer requires years of training or a specialized team. That barrier, which until now protected many organizations without them even knowing it, is starting to disappear," explains the team at Jiratek, a company specializing in cybersecurity, to Vozpópuli.
If a system like Mythos were compromised, the damage could be fatal. Cybersecurity experts explain that the possibilities include someone stealing the model or a person with legitimate access using it maliciously. In either case, "the most immediate damage is having a map of unknown vulnerabilities before anyone has patched them. A clean attack window, with no defenses prepared." However, they add that "the most serious scenario is not stealing data, but taking control of physical systems."
"A company that suffers such a breach doesn't just lose data: it loses the trust of customers and regulators, with losses of up to 50% of turnover. If the flaw was known and not corrected, legal exposure is added on top, with fines of up to 4% of the company's annual turnover," the Jiratek team explains.
To mitigate these risks, the company led by Dario Amodei has opted to select a group of companies—mostly American, including Apple, Microsoft, and Amazon—to help build a defensive barrier before hackers can bypass them.
Anthropic has dubbed this group of one hundred companies Project Glasswing, which works to strengthen the security of what they describe as "the most critical software in the world." One UK-based company has already claimed that Mythos has the ability to carry out complex cyberattacks that no previous AI system had ever achieved.
"The most exposed sector is the industrial one: energy, utilities, water, transport, manufacturing. Infrastructure with decades of age, systems that cannot be easily patched, and an investment in cybersecurity that, in our experience, remains insufficient. That is where the impact could be most severe," they emphasize.
Mythos is an AI system similar to Anthropic's Claude, OpenAI's ChatGPT, or Google's Gemini. However, the difference here is that it is much more powerful than its siblings. According to Anthropic members, the AI system "has identified thousands of additional high and critical severity vulnerabilities." They also added that they are "working to disclose them responsibly."
According to the company's blog, during testing, the AI was able to find unknown security flaws (zero-days) and also analyze and exploit other known flaws that were not yet fully patched. "More than 99% of the vulnerabilities we have detected have not yet been fixed, so it would be irresponsible to disclose details about them," they wrote on April 7.
"However, even the 1% of bugs we can mention offers a clear view of a significant leap in what we consider the next generation of model cybersecurity capabilities—an advance that justifies coordinated and substantial defensive action across the industry," they added.
Should this type of system be used with malicious intent, users would have access to the computer systems of organizations and companies, critical infrastructure such as hospitals or power grids, government security systems, and the compromise of operating systems and browsers.
International Alert
Alarms have gone off internationally, triggering fear in many countries. The United Kingdom is one of the countries where this fear has prompted the government to force its leading companies to strengthen security systems. Meanwhile, the European Commission is examining the potential risks in the field of computer security and has held talks with Anthropic. However, the reality is that, for now, they have not been granted the opportunity to directly investigate the threats posed by the model.
In the United States, just a few weeks ago, Treasury Secretary Scott Bessent and Federal Reserve Chairman Jerome Powell held several meetings with top executives from Morgan Stanley, Wells Fargo, Bank of America, Citigroup, and Goldman Sachs to address the issue. JPMorgan Chase CEO Jamie Dimon was also invited but was ultimately unable to attend.
Last February, the U.S. Department of Defense, led by Pete Hegseth, labeled Anthropic a threat to national security. Mythos has also reached the White House. Anthropic owner Dario Amodei visited a week ago to discuss Mythos and its scope. Just a few days later, President Donald Trump stated the following in an interview with CNBC: "We had very productive conversations with them, and I think they are making progress." "They are very smart, and I think they can be of great use," he added.
Source: Voz Populi