Dario Amodei, CEO of Anthropic, an American artificial intelligence company founded by former members of OpenAI and creator of ChatGPT, publicly admitted that Mythos-class artificial intelligence models represent a real risk to the financial sector, critical infrastructure and national security.
The statement appears in the essay “Policy on the AI Exponential”, published this June 10 on his personal website, and occurs a day after the company launched Claude Fable 5 with explicit cybersecurity restrictions.
According to Amodei, Mythos Preview has already altered the global cybersecurity landscape and constitutes evidence that border models are tools of national strategic consequence. The executive maintains that these risks are not hypothetical: They are a present reality that requires a binding regulatory response..
It must be considered, in this sense, that Anthropic develops the Claude models and positions itself as one of the companies in the sector with the greatest influence in the global regulatory debate. Something that gives direct weight to the warnings of its CEO.
The immediate context is relevant. As reported by CriptoNoticias, through Project Glasswing – the restricted program with which Anthropic distributed Mythos Preview since April – the company and its partners identified more than 10,000 high or critical severity vulnerabilities in global systemic infrastructure softwarel. Cloudflare, Mozilla, Microsoft, Oracle and Palo Alto Networks were included, as well as banks and critical infrastructure institutions.
Companies must self-regulate
Amodei suggests that models that exceed a certain computational threshold must undergo mandatory audits by third parties in four areas:
- Cybersecurity
- Biological weapons.
- Loss of control of AI and research systems
- Automated development that accelerates these risks.
According to the essay, the government should have the power to block or reverse the deployment of a model if it represents an unacceptable risk in any of those categories.
The firm proposes to model AI regulation on the United States Federal Aviation Administration (FAA): just as airplanes must pass technical tests before operating, frontier models should require certification prior to launch.
The development companies, according to the test, must also protect the weights of their models, carry out periodic penetration tests and immediately report any critical incident to the authorities.
The CEO recognizes the paradox implicit in his position: Anthropic builds the technology that he himself describes as dangerousand is now promoting its binding regulation, after having supported only voluntary transparency frameworks until 2025.
On a geopolitical scale, Amodei argues that democracies must coordinate AI-based cyber defenses and share intelligence to stop malicious uses. According to the essay, a nation with advanced AI versus one without it would be equivalent to a modern army versus a medieval one.
Amodei warns that the cyber risks that Mythos Preview highlighted will not be the last: according to its analysis, biological risks derived from next-generation models could materialize in the short term. He adds that the risks that Mythos Preview exposed in cybersecurity they do not represent the arrival point, but rather the starting point.
This is a controversial position, as critics argue that AI CEOs’ warnings also serve to justify regulations that favor them and slow down open innovation. In the cryptocurrency ecosystem, where AI is already used for both contract audits and advanced exploits, the debate is open.
In this framework, the regulation proposed today is not designed for the most serious problem it anticipates, but for the one that is already here, with the explicit warning that the regulatory apparatus It will have to scale as fast as the technology does..