It all began with the launch of new AI models by Anthropic, the company behind the Claude chatbot. The models unveiled were Fable 5 and Mythos 5, and they boasted the status of the company's most powerful AI models to date. But within a few days, the whole matter turned into a major controversy involving not only the US government but also the National Security Agency (NSA).
What actually happened
Anthropic launched Fable 5 and Mythos 5 at the beginning of last month. Just three days later, however, according to Anthropic, Amazon researchers found a way to bypass the safeguards of the first model mentioned. This led to the discovery that the model was capable of helping identify software vulnerabilities.
However, the issue was not merely that the new AI model could help uncover vulnerabilities, but above all the risk that it could become a dangerous tool in the hands of someone seeking to attack a system.
And that was far from all. According to reports, the second AI model, Mythos, managed to penetrate nearly all of the NSA's classified systems within just a few hours. This claim was made during a Senate hearing and was subsequently reported by The Economist.
Immediate shutdown of the models
According to available information, the US government did not want to risk these capabilities falling into the hands of individuals or organizations outside the approved circle. The administration therefore ordered Anthropic to restrict access to the Fable 5 and Mythos 5 models exclusively to US citizens. However, verifying every user's citizenship in real time was not feasible, so Anthropic chose a different solution and shut down both models globally.
AP subsequently reported that Fable 5 was made available again after modifications. Mythos 5, however, remained available only to selected US organizations approved by the government.
Mythos and the NSA
According to The Economist, Senator Mark Warner stated during the hearing that, according to the head of the NSA, Mythos had penetrated nearly all classified systems within a few hours, which appeared extremely alarming. However, details about the entire affair were lacking, and it was therefore unclear whether this was a test in a controlled environment, an internal assessment, or something else. This made the subsequent radical solution of shutting down the models all the more surprising.
Anthropic's response
Anthropic said that it had worked with the government and Amazon to develop a new security measure. Specifically, it is a classifier designed to detect suspicious requests and block attacks in more than 99% of cases.
The company also announced closer cooperation with the US government and other major players, including Amazon, Microsoft, and Google. The goal is to improve model testing before launch and standardize the criteria for assessing similar attacks.
AI as a good servant but a bad master
Attempts to bypass the security rules of AI models are entirely commonplace. They are known as jailbreaks, and researchers test them constantly. This makes the government's swift and forceful response, which led to restricted access to the newly launched models, all the more surprising.
The entire controversy demonstrates the direction in which the AI trend in cybersecurity is heading. Artificial intelligence is no longer merely a smart assistant that checks code and makes work easier. For many countries, the most powerful models are becoming a strategic tool. Yet they may not only protect systems; in the wrong hands, they can become a dangerous means of gaining access.
And that is precisely why Mythos and Fable are not merely two more names on the long list of AI developments. They are a reminder that as models become more capable, the debate will increasingly focus not only on what they can do, but above all on who should have access to them.
Source: Security Affairs, The Economist, Anthropic, AP News, TechSpot
