TLDR: Anthropic has introduced a new, targeted transparency framework aimed at enhancing safety and accountability for the development of frontier AI systems. The framework proposes mandatory disclosure requirements for the largest AI model developers, focusing on secure development practices, risk mitigation, and public reporting, while exempting smaller entities to foster innovation.
San Francisco, CA – In a significant move to address growing concerns surrounding the rapid advancement of artificial intelligence, Anthropic, a leading AI safety company, announced on July 7, 2025, a comprehensive and targeted transparency framework for frontier AI systems. This proposal aims to establish clear disclosure requirements for safety practices among the most powerful AI model developers, ensuring greater public safety and accountability.
The framework is specifically designed to apply only to the largest AI model developers, distinguished by substantial financial thresholds. Companies would be covered if they meet annual revenue cutoffs on the order of $100 million, or have R&D or capital expenditures on the order of $1 billion annually. This targeted approach deliberately excludes smaller developers and startups, aiming to avoid burdening the nascent AI ecosystem and those developing models at lower risk of national security implications or catastrophic harm.
A core tenet of Anthropic’s proposal is the requirement for covered AI companies to develop and adhere to ‘Secure Development Frameworks‘ (SDFs) prior to the deployment of any new model. These SDFs must detail how companies assess and mitigate ‘Catastrophic Risks,’ which are defined to include potential harms such as those related to Chemical, Biological, Radiological, and Nuclear (CBRN) threats, or models acting autonomously in ways contrary to developer intent.
Minimum standards for these SDFs include identifying the models they apply to, describing assessment and mitigation approaches for catastrophic risks, outlining processes for modifying the SDF, identifying a responsible corporate officer for compliance, and establishing robust whistleblower processes for employees to report safety concerns without fear of retaliation. Companies would also be required to confirm implementation of their SDFs before model deployment and retain copies for at least five years.
Beyond pre-deployment requirements, the framework mandates minimum transparency requirements. Covered companies must publicly disclose their SDFs on a readily accessible, public-facing website. Furthermore, at the time of deployment of a new model or a substantial new capability, they must publish a ‘system card‘ or similar documentation. This documentation should summarize model testing and evaluation procedures, results, and any required mitigations under the SDF. Companies must also certify compliance with SDF requirements and disclose this on a public website. While the framework allows for redaction of trade secrets or information that would compromise public safety or model security, any such omissions must be briefly identified and justified.
Also Read:
- Proactive Vulnerability Testing: The Rise of AI Red Teaming for Enhanced Safety
- Singapore Unveils Global AI Assurance Sandbox to Enhance Real-World Testing and Governance
Anthropic’s initiative comes at a critical juncture as the AI industry faces increasing scrutiny over safety, bias, and societal risks. The company states that the framework aims to provide policymakers with the evidence needed to determine whether additional regulation is necessary, while also giving the public vital information about the technology. This approach seeks to balance the need for transparency with the agility required for private sector innovation, ensuring that AI’s transformative potential, from drug discovery to national security, can be realized responsibly. The framework is flexible, allowing for implementation at federal, state, or international levels, and is poised to influence regulatory approaches and best practices for leading AI companies worldwide.
