A framework for AI development transparency

Source: Anthropic Date: 2025-07-07 URL: https://www.anthropic.com/news/the-need-for-transparency-in-frontier-ai

Summary

Anthropic published a targeted AI transparency framework (July 2025) for federal, state, or international application, scoped to the largest AI developers. Core recommendations: mandatory secure development framework documentation for CBRN/autonomy risks; public disclosure of risk assessment frameworks (with redactions); system cards at deployment and after substantial revisions; prohibition on false compliance statements (whistleblower enablement); threshold: $100M+ annual revenue or $1B+ in R&D/capex. Framework explicitly avoids heavy prescription, citing rapid evaluation method obsolescence.

Implications

• Policy / Anthropic as regulatory architect. The $100M+/$1B+ threshold is calibrated to cover every frontier lab (OpenAI, Google DeepMind, Meta, Anthropic) while excluding most startups. Anthropic is proposing a compliance structure that applies to exactly its competitive set — a smart regulatory strategy if adopted.
• CBRN + autonomy as named risk categories. Scoping mandatory disclosure specifically to CBRN and autonomy-related harms maps directly to the ASL-3 trigger criteria (May 2025). The framework is the policy version of the RSP’s capability-based thresholds.
• “Evaluation methods become outdated within months.” Building anti-rigidity into the framework explicitly is the mechanism that allows Anthropic’s own RSP updates to remain compliant even as evaluation methods evolve. It’s both honest and self-serving.
• System cards as minimal transparency. Requiring system cards at deployment (not capability evaluation scores) is the minimum meaningful disclosure — it’s lower overhead than Anthropic’s own RSP disclosures, suggesting this is a floor proposal, not a ceiling.
• Watch: whether the framework influenced US AI legislation or EU AI Act implementation; whether the $100M threshold was adopted as a regulatory trigger; how system card requirements evolved relative to Anthropic’s actual published system cards.