Anthropic proposes new transparency framework for frontier AI models

Anthropic announced on July 7, 2025, a comprehensive transparency framework designed to address growing concerns about frontier artificial intelligence development. The proposal targets the largest AI model developers through specific revenue and spending thresholds while establishing mandatory disclosure requirements for safety practices.

According to the announcement, the framework would apply only to companies meeting substantial financial criteria. "Annual revenue cutoff amounts on the order of $100 million; or R&D or capital expenditures on the order of $1 billion annually" represent the proposed thresholds for coverage under the transparency requirements.

Summary

Who: Anthropic, an AI safety company, proposed the transparency framework targeting frontier AI model developers meeting specific revenue ($100 million annually) or spending ($1 billion in R&D/capital expenditures) thresholds.

What: A comprehensive transparency framework requiring covered companies to publicly disclose Secure Development Frameworks detailing safety practices, publish system cards summarizing testing procedures, and implement whistleblower protections for employees reporting safety concerns.

When: Announced July 7, 2025, with implementation proposed at federal, state, or international levels. The framework would apply only after jurisdictions adopt the requirements through legislation or regulation.

Where: Proposed for implementation in any jurisdiction, with flexibility for federal, state, or international adoption. The framework would cover companies meeting thresholds regardless of location when operating in adopting jurisdictions.

Why: Growing concerns about frontier AI development occurring without adequate transparency or accountability, particularly regarding catastrophic risks including chemical, biological, radiological, and nuclear harms, plus risks from misaligned autonomous AI systems acting contrary to developer intentions.

The framework centers on requiring covered companies to develop and publicly disclose Secure Development Frameworks (SDFs). These documents must detail how companies assess and mitigate catastrophic risks from their AI models. Catastrophic risks specifically include "the creation of chemical, biological, radiological and nuclear harms, as well as harms caused by misaligned model autonomy."

Companies subject to the framework would face several mandatory transparency requirements. They must publish their Secure Development Framework on publicly accessible websites registered to their organization. Additionally, they must provide system cards summarizing testing procedures, evaluation results, and implemented mitigations at the time of model deployment.

The proposal includes provisions for protecting sensitive information. Companies may redact trade secrets, confidential business information, or details that could compromise public safety or model security. However, any redactions must be briefly identified and justified in the public version.

Enforcement mechanisms focus on preventing false statements about compliance rather than prescriptive regulatory oversight. The framework explicitly prohibits "intentionally false or materially misleading statements related to SDF compliance" and authorizes attorneys general to seek civil penalties for material violations. Companies receive a 30-day right to cure violations before penalties take effect.

The timing reflects growing policy attention to AI governance. Recent developments include the OECD Expert Group's comprehensive framework addressing AI development through 2040, which emphasized mandatory disclosure requirements for high-risk AI systems. German data protection authorities also published detailed technical requirements for AI systems in June 2025, establishing lifecycle-based governance approaches.

The proposed framework deliberately avoids prescriptive technical standards. "We recognize that as the science of AI continues to evolve, any regulatory effort must remain lightweight and flexible," according to the announcement. This approach acknowledges the rapid pace of technological change where evaluation methods become outdated within months.

The framework builds on existing industry practices. Leading companies including Anthropic, Google DeepMind, OpenAI, and Microsoft have already implemented similar approaches through voluntary Responsible Scaling Policies. "Putting a Secure Development Framework transparency requirement into law would not only standardize industry best practices without setting them in stone, it would also ensure that the disclosures (which are now voluntary) could not be withdrawn in the future as models become more powerful."

Whistleblower protections receive specific attention in the proposal. Secure Development Frameworks must describe processes for employees to raise concerns about framework content and implementation, along with protections from retaliation. This provision creates legal pathways for internal oversight of safety practices.

The framework's scope deliberately excludes smaller developers and startups. According to the proposal, transparency requirements should not burden "the startup ecosystem and small developers with models at low risk to national security or for causing catastrophic harm." The revenue and spending thresholds aim to focus oversight on companies with sufficient resources to develop truly frontier models.

Implementation would occur at federal, state, or international levels according to the proposal. The flexible jurisdictional approach recognizes varying regulatory environments while maintaining consistent core requirements for transparency and accountability.

Industry context shows mixed approaches to AI transparency. Digital advertising platforms have increasingly implemented ad libraries providing unprecedented visibility into advertising practices. However, research indicates 80% of companies actively block AI language models from accessing their websites, suggesting ongoing tensions between transparency and competitive interests.

The proposal acknowledges differing views on catastrophic AI risks. "Views differ on whether and when AI models could pose catastrophic risks," the announcement states. The transparency framework aims to provide policymakers with evidence needed to determine whether additional regulation proves necessary while giving the public important information about the technology.

Technical requirements for Secure Development Frameworks include specific elements. Companies must identify which models the framework covers, describe assessment and mitigation approaches for catastrophic risks, address modification processes, identify responsible corporate officers, and establish employee reporting mechanisms. Frameworks must be retained for at least five years.

System card requirements mandate documentation of testing procedures, evaluation results, and required mitigations. These documents must be published at deployment and updated when models receive substantial capability additions. The requirements parallel existing practices at leading AI laboratories while creating legal obligations for disclosure.

The proposal emphasizes balancing innovation with safety considerations. "Without safe and responsible development, a single catastrophic failure could halt progress for decades," according to the announcement. The framework positions transparency as enabling continued innovation while providing public oversight of safety practices.

Regulatory momentum continues building across jurisdictions. European authorities have implemented comprehensive AI governance through the EU AI Act, while individual countries develop national frameworks. Apple faced significant regulatory challenges over its App Tracking Transparency framework, demonstrating how transparency requirements can create competition law issues when implemented by dominant platforms.

The framework's lightweight approach contrasts with more prescriptive regulatory proposals. Rather than mandating specific technical approaches, it focuses on disclosure requirements that enable external oversight while preserving company flexibility in implementation methods. This structure aims to avoid stifling innovation while ensuring accountability for safety practices.

Companies meeting the thresholds would need to adapt existing practices to meet formal disclosure requirements. Many leading AI developers already publish safety documentation voluntarily, suggesting implementation costs may remain manageable for covered companies. However, the legal obligations create permanent commitments that companies cannot withdraw as competitive pressures increase.

The announcement comes as frontier AI capabilities continue advancing rapidly. Recent model releases demonstrate increasing sophistication in reasoning, coding, and multimodal processing. These advances underscore the importance of transparency frameworks that can evolve with technological progress while maintaining consistent oversight principles.

Public safety considerations drive the framework's focus on catastrophic risks. Chemical, biological, radiological, and nuclear threats represent areas where AI capabilities could enable significant harm if deployed maliciously or developed without appropriate safeguards. The autonomous action category addresses concerns about models that might act contrary to user intentions or developer goals.

The proposal's emphasis on industry flexibility reflects ongoing uncertainty about optimal safety practices. "Secure Development Frameworks are still an evolving safety tool, so any proposal should strive for flexibility," according to the announcement. This acknowledgment suggests the framework would adapt as safety science advances and best practices emerge.

Private sector agility receives protection through the framework's design. Rather than government-imposed technical standards, the approach relies on company-developed frameworks subject to public disclosure. This structure preserves innovation incentives while creating accountability mechanisms for safety practices and risk mitigation approaches.

The framework addresses growing public concern about AI development occurring behind closed doors. Transparency requirements enable researchers, policymakers, and the public to understand current AI capabilities and safety practices. This visibility could help build public trust while enabling more informed policy decisions about additional oversight needs.

Timeline

• July 7, 2025: Anthropic announces comprehensive transparency framework for frontier AI development
• June 2025: German authorities issue comprehensive AI development guidelines covering entire development lifecycle
• May 14, 2025: Belgian court limits IAB Europe's role in TCF framework addressing transparency in digital advertising
• March 30, 2025: Apple fined €150 million for anticompetitive App Tracking Transparency framework
• March 27, 2025: Digital ad libraries address industry transparency demands across major platforms
• February 13, 2025: German regulator finds Apple's app tracking rules may violate competition law
• November 2024: OECD expert group maps future AI landscape through 2040