According to Odaily, Anthropic CEO Dario Amodei has taken one of his strongest public positions to date in favor of tighter artificial intelligence regulation. In a long post published on Wednesday, Amodei said governments should have the authority to stop the deployment of new AI models when specific risks are identified. He argued that AI models should be subject to mandatory third-party testing, with the goal of assessing risks across multiple areas before deployment decisions are made.
Mandatory External Testing for AI Models
Amodei wrote that if an AI system is determined to pose “unacceptable risks,” then “the government should have the power to prevent or restrain its deployment.” His statement shifts the emphasis beyond internal company review and toward a formal external testing process. Under that approach, new models would be evaluated by third parties before being released more broadly, especially when their capabilities raise concerns in sensitive domains.
To explain the type of oversight he believes is appropriate, Amodei compared the current stage of AI development with regulated technologies such as cars, aircraft and medicines. He wrote: “I think, at least during the current period of exponential growth, the right analogy is to cars, airplanes, or drugs — technologies that are critical to the modern economy, but can cause large numbers of deaths if designed or used improperly.” The comparison frames AI not only as an important technology for the economy, but also as a category where deployment safety and misuse risks should be assessed before broad release.
Anthropic’s Mythos Case
The comments come after Anthropic previously warned that its AI model Mythos was capable of discovering and exploiting critical software vulnerabilities. Because of those capabilities, the company decided to make the model available only to a small number of partners rather than release it more widely. That earlier decision provides a concrete example of the type of deployment restraint referenced in Amodei’s latest argument.
This week, Anthropic also released a new version that removed the relevant cyberattack capabilities. In Amodei’s formulation, the central issue is the threshold of “unacceptable risks.” If a model is found to cross that threshold, he argues, the government should be able to block or limit its deployment. His remarks add a more explicit regulatory dimension to Anthropic’s prior handling of Mythos and its decision to restrict access to a model with sensitive cybersecurity functions.