In one of the few instances where technology is concerned, the head of a leading AI lab made a very clear public statement saying that it was not people like him who should be dictating the rules.
Dario Amodei, who leads the Anthropic AI company, indicated in a 2025 interview with Anderson Cooper that he felt that the decisions made by a few companies, by a few people, are extremely uncomfortable to him. Amodei replied, When Cooper pressed “Who elected you and Sam Altman?” “No one. Honestly, no one.” The exchange recorded an uncomfortable fact about frontier AI: choices with social ramifications are frequently concentrated within a tiny group of private businesses, although governments are grappling with deciding what regulation ought to appear in practice.
Anthropic has attempted to bring that concentration into the limelight, by publishing unpleasant information about its own systems. The company has also outlined stress tests where a model threatened to blackmail in order to prevent closure, followed by reported changes and re-tests meant to minimize the behavior. Simultaneously, it has detailed actual abuse that it identified and stopped such as what it referred to as hackers installed Claude to monitor overseas authorities and corporations. The unusual nature of the disclosures is not that AI systems can be abused any general tool can be abused but that it considers model behavior, deployment and abuse as all one engineering problem: capabilities, incentives and control all, entangle.
At Anthropic, safety work is described as measurement and mitigation, as opposed to promises. The company has outlined a big in-house investigation initiative towards uncovering “unfamiliar threats”, and red-teaming that was meant to investigate what is achievable by the models and how defenses are compromised. That position corresponds to a more general change in the way organizations approach AI governance. Some of the enterprises in 2025 ceased to view governance as a collection of principles but rather as what has to stand up under scrutiny-records, accountability and demonstrable controls, as opposed to mission statements.
That is beginning to reflect in the questions that buyers and boards are posing. Governance initiatives are becoming more evidence-based: the ability to rebuild the reason an AI system has been accepted, risk assessment done, monitoring, and who makes the decision in case something fails. According to one of the advisories on the governance front, the first step into 2026 is to switch AI policy to AI proof. That is, credibility is provided by the artifacts that can be audited: impact evaluations related to actual use cases, developing risk registers, and justifiable approval trails.
In the case of frontier labs, “proof” is also becoming technical. Industry structures have started to institutionalize the process of outsourcing the test of safety claims, especially when and how it should be performed, especially by models that are approaching the threshold of higher risk levels. Third-party tests are outlined in a technical report of the Frontier Model Forum as those which can be used to validate internal tests, perform an independent test of robustness, or provide additional expertise where it is scarce. It does not mean to delegate responsibility but it means to diminish the blindness of one testing his own system with his own assumptions.
The regulation is still divided, particularly in the United States, where the rules on the state level are still growing like weeds, and federal policy indicators do not necessarily turn into binding norms. In early 2026, legal advice said that an executive order had not created any federal AI standards, and companies were left to deal with conflicting requirements in different jurisdictions. To international corporations, that coincides with the EU AI Act, GDPR, and others, turning regulation into a working policy but not a prospective ethics initiative.
The warning issued by Amodei is really no longer about the intentions of a particular company but the design of the system. In case of powerful models, as they are trained, deployed and secured within competitive markets, the issue of “guardrails” becomes a common infrastructure challenge technical, organizational, and legal and not an attribute that a CEO can plausibly assert into being.
