Published: July 7, 2025 | Category: Policy
A new proposal aims to increase public accountability for the companies building the most capable "frontier" AI systems. The authors argue that while industry, governments and academia are working toward shared safety standards and comprehensive evaluation methods — a process that could take months or years — interim measures are needed to ensure powerful AI is developed securely, responsibly and transparently.
Purpose and nature of the proposal
The recommendation sets out a targeted transparency framework that could be implemented at the federal, state or international level and that would apply only to the largest developers of frontier models. The approach intentionally avoids heavy prescriptiveness: regulators should keep requirements lightweight and flexible so they do not impede innovation or slow the realization of AI benefits, such as life‑saving drug discovery, faster delivery of public benefits, and critical national security functions.
The authors note that rigid government standards can be counterproductive because evaluation methods can become outdated within months given the pace of technological change.
Who should be covered and why
Key elements of the framework include:
-
Limit scope to the largest frontier model developers: "Frontier" models would be distinguished using combined thresholds — for example, compute power, compute cost, evaluation performance, annual revenue and R&D spend — so that small developers and startups that pose low risk to national security or catastrophic harm are not unduly burdened.
-
Provide appropriate exemptions: the framework would include exemptions for smaller developers, and the proposal solicits input from the startup community on where to set thresholds.
-
Illustrative thresholds: the document offers example cutoffs such as an annual revenue threshold on the order of $100 million, or annual R&D or capital expenditures on the order of $1 billion. The authors recommend periodically reviewing these scoping thresholds as the technology and industry evolve.
Secure Development Framework
Covered frontier developers would be required to maintain a Secure Development Framework that explains how they assess and mitigate unreasonable risks posed by a model. These risks explicitly include the potential creation of chemical, biological, radiological and nuclear harms, as well as harms from misaligned model autonomy.
Because Secure Development Frameworks are an evolving safety tool, any regulatory requirement should aim for flexibility.
Public disclosure and self‑certification
The Secure Development Framework should be disclosed publicly on a company‑maintained, registered website, with reasonable redactions for sensitive information. The public disclosure should be accompanied by a self‑certification stating that the lab is complying with the published framework.
Making these frameworks public enables researchers, governments and the public to stay informed about the safety practices used for models that are deployed today.
System cards
The framework also recommends publishing system cards or similar documentation that summarize testing and evaluation procedures, results and required mitigations, again with appropriate redactions where details could compromise public safety or model security. System cards should be disclosed at deployment and updated if the model is substantially revised.
Whistleblower protection and prohibiting false statements
The proposal explicitly makes it a legal violation for a lab to lie about its compliance with its Secure Development Framework. This clarification creates a clear legal basis for whistleblower protections to apply and directs enforcement at labs that have intentionally misrepresented their compliance.
Minimum, evolving standards
A practical transparency framework should set a minimum set of standards that enhance security and public safety while accommodating the evolving nature of AI development. Because AI safety and security practices are still nascent and frontier developers such as Anthropic are actively researching best practices, requirements should begin as flexible, lightweight obligations that can adapt as consensus best practices emerge among industry, government and other stakeholders.
The approach aims to illuminate industry best practices for safety and help establish a baseline for how responsible labs train models, enabling the public and policymakers to distinguish responsible from irresponsible practices. The proposal notes that Secure Development Frameworks are similar in spirit to existing policies such as Anthropic’s Responsible Scaling Policy and practices already used by Google DeepMind, OpenAI and Microsoft when releasing frontier models.
Why this matters now
Applying a Secure Development Framework requirement in law would standardize disclosure of industry practices without locking those practices in permanently, and would prevent voluntary disclosures from being withdrawn later as models become more powerful. Views differ on if and when AI could pose catastrophic risks, but transparency requirements for Secure Development Frameworks and system cards could provide policymakers with the evidence needed to decide whether further regulation is warranted and give the public important information about powerful new technology.
Without safe and responsible development, the authors warn, a single catastrophic failure could set back progress for decades. The proposed transparency framework is presented as a practical first step: it provides public visibility into safety practices while preserving private sector agility to deliver AI’s transformative potential.
