Abstract:
Anthropic CEO Dario Amodei called for controlling the speed of improvement in the capabilities of cutting-edge artificial intelligence models to buy time for security assessments and protective measures. He proposed introducing permanent independent assessors, coordinating safety standards of major developers, and promoting cross-border risk management and control cooperation.
Independent evaluators gain internal access
Amodei proposed that cutting-edge AI companies should provide continuous, near-employee-level access to qualified independent assessment teams so that they can review model testing, risk reporting, safety frameworks and incident handling processes.
Anthropic is committed to being the first to implement this measure. Outside personnel will have access similar to that of the company's risk assessment team and be able to publish conclusions without Anthropic's editorial control.
This arrangement still needs to resolve issues such as the selection of evaluators, funding sources and confidentiality obligations. Model weights, training infrastructure, and security vulnerabilities are highly sensitive information, and expanding access rights will also increase information security requirements.
Industry coordination limits unprotected competency competition
The second recommendation calls for major AI labs to develop common standards for risk testing, safety thresholds, and model release conditions. Amodei believes that competitive pressure may prompt companies to speed up training and release pace, making it difficult for protective measures to keep up with model capabilities in a timely manner.
Anthropic's Responsible Scaling Policy has adopted a tiered mechanism. When models cross biological, cyber, automated R&D, or loss of control risk thresholds, companies need to increase their deployment protection and information security levels.
Inter-enterprise cooperation may involve antitrust boundaries. There is a public interest in working together to develop safety testing standards, but arrangements may still be subject to regulatory scrutiny if they restrict smaller players from entering the market, coordinate product launches or reduce competition.
International cooperation covers cross-border risks
The third recommendation involves intergovernmental cooperation, including establishing minimum security arrangements with strategic competitors such as China. Amodei believes that models and technologies can flow across borders, and slowing down the development speed of a single country or enterprise may lead to the transfer of high-risk R&D to regions with weaker regulations.
Potential risks listed by Anthropic include using models to assist bioweapons activities, large-scale discovery of software vulnerabilities, attacks on critical infrastructure, and AI systems breaking out of developer control. AI automatically participates in the development of next-generation models, which may also further accelerate capability improvements.
The company proposes that cutting-edge developers should publish system cards and regular risk reports, accept independent review, and provide necessary information to designated government agencies. Governments should have limited powers to prevent or delay deployment of models that could cause catastrophic damage, with clear thresholds and appeals mechanisms in place.
Industry support has not yet been translated into an agreement
OpenAI CEO Sam Altman supports giving independent evaluators internal access and said OpenAI will adopt similar arrangements. Google DeepMind CEO Demis Hassabis and xAI head Elon Musk have also publicly supported Amodei’s overall direction.
Relevant statements have not yet formed a binding industry agreement. Which capability indicators will trigger slowdowns or suspensions, whether independent reviews will have mandatory effect, and how cross-border agreements will be enforced remain to be further determined by businesses and governments.
Comments