The Future of AI Safety - Anthropic's Transparency Report

Artificial intelligence is advancing rapidly, raising important questions about safety, risk, and governance. To address these concerns, the U.S. State of California introduced the Transparency in Frontier Artificial Intelligence Act (the "TFAIA", "CA SB 53" or the "Act") in 2025, which requires large frontier model AI developers (annual gross revenues in excess of 500 million USD) to create and publish comprehensive Frontier AI Frameworks and Transparency Reports.

This series examines how three frontier‑model developers—Anthropic, OpenAI, and Google DeepMind—are meeting TFAIA requirements. This post focuses on Anthropic’s transparency reports, with upcoming entries covering OpenAI and Google DeepMind and concluding with a comparative analysis. Reviewing these reports and frameworks shows how frontier AI teams are advancing safer and more transparent systems.

Our previous blog post on the TFAIA points out that while the TFAIA is focused on large developers, smaller developers and SMBs should take note - the TFAIA acknowledges that foundation models developed by smaller companies may eventually pose risks that warrant additional regulation, and the Act is poised to serve as a template for future legislation in California and other U.S. states.

Transparency Reports

Let's begin with an overview of transparency reports, as the TFAIA (CA SB 53) establishes comparatively straightforward requirements for these reports (relative to the requirements for Frontier AI Frameworks).

Transparency Reports. CA SB 53 requires that, when deploying a new frontier model or a substantially modified version, developers must publish a clear and conspicuous transparency report on their websites containing the following (which may be incorporated into a system card, model card or similar document):

• Basic Info: (a) internet website of the frontier developer; (b) contact information; (c) model release date; (d) model-supported languages; (e) supported output-modalities; (f) intended uses of the model; (g) and general use restrictions.

  
  

• Compliance Summaries: (a) catastrophic risk assessment results, (b) the extent of third party evaluator involvement and (c) other steps taken to fulfill the requirements of the frontier AI framework.

With these statutory requirements in mind, the following review assesses how Anthropic’s Transparency Hub measures up to the TFAIA’s transparency‑reporting mandate.

Anthropic's Transparency Hub

Anthropic's website, includes a "Commitments" drop down menu with initiatives for Transparency, a Responsible Scaling Policy and a Trust Center for Security and Compliance. These policies are easy to locate and navigate, reflecting strong attention to search-ability and ease of access on Anthropic’s part.

Model Reports, System Card

The Transparency Hub provides a comprehensive set of disclosures that map closely to the TFAIA’s transparency‑reporting requirements. It includes model reports for all major Claude releases, along with system documentation, trust and security materials, and key voluntary commitments. The model report page allows selection of various models with summaries of essential information, condensing key details about the models' capabilities, safety evaluations, and deployment safeguards. The page also includes a link to the Claude Opus 4.5 System Card.

The checklist below focuses specifically on how the Transparency Hub materials for Claude Opus 4.5 align with the TFAIA’s transparency‑reporting mandate.

Claude Opus 4.5

Information Required	Included?	Location	Notes
Frontier developer website	√	System Card	This requirement is somewhat redundant, as the report must appear on the developer’s website. Anthropic nonetheless addresses it explicitly by listing anthropic.com on the first page of its system card.
Contact information (mechanism that enables a natural person to communicate with the frontier developer)	√	Support Center ; Support Page ; Privacy and Legal ; Online Safety Contacts; EU Designated Point of Contact	Various mechanisms for communication exist across the broader site but this is an area where Anthropic could streamline access and link information directly to the Transparency Hub. The site includes a support center with guidance for paid and free users, a chatbot, and articles covering privacy and legal information with contact forms.
Model release date	√	Model Report	The model report includes a release month and year (November 2025 for Claude Opus 4.5).
Supported languages	√	Model Report; System Card	The bill’s text is vague on whether it refers to programming languages or natural languages, but Anthropic addresses both. Its model report and system card note that Claude Opus 4.5 expands single‑turn evaluations to English, Arabic, French, Korean, Mandarin Chinese, and Russian, and the system card also cites SWE‑bench results showing the model can operate across nine programming languages.
Supported output modalities	√	Model Report	The model report includes modalities, noting that Claude Opus 4.5 can understand both text (including voice dictation) and image inputs, engaging in conversation, analysis, coding, and creative tasks. Claude can output text, including text-based artifacts, diagrams, and audio via text-to-speech.
Intended uses of the model	√	Model Report; System Card	The bill is also vague on “intended use,” given that frontier models can be applied in highly flexible ways. Anthropic provides brief use descriptions, access surfaces and software integration guidance in its model cards, and the Claude Opus 4.5 system card adds a capabilities‑evaluation section that was not included in earlier releases.
General use restrictions	√	Model Report	The model report includes Acceptable Uses and a link to their Usage Policy.
Catastrophic risk assessment results	√	Model Report; System Card	The model report and system card present results from various catastrophic risk assessments, including Anthropic’s “Responsible Scaling Policy” evaluation, CBRN evaluations, biological‑risk assessments, and capability threshold testing (could the model automate an entry level position or junior position). They also summarize evaluations of harmlessness, agentic safety, behavioral risks, evaluation awareness, and cybersecurity.
3rd party evaluator involvement	√	Voluntary Commitments; Trust Center	Anthropic clearly notes on the "voluntary commitments" page, that they engage with independent assessors to evaluate the effectiveness of their cybersecurity measures and participate in third-party test and evaluation schemes periodically, with compliance artifacts available on their "trust center" page.
Compliance with Frontier AI Framework	√	Model Report; System Card	Because the model report and system card disclose results across catastrophic‑risk evaluations, capability‑threshold testing, and assessments of harmlessness, agentic safety, behavioral risks, evaluation awareness, and cybersecurity, Anthropic is effectively demonstrating compliance with its frontier‑AI framework. That said, more to come on specific compliance in the next installment of this blog series.

How did Anthropic fare? It turns out very well. Anthropic’s Transparency Hub provides a clear, well‑organized set of disclosures that align closely with the TFAIA’s transparency reporting mandate. The core requirements are met comprehensively. Overall, Anthropic presents an excellent example of implementing the new requirements. No doubt that as the TFAIA continues to shape baseline expectations for transparency and risk governance across the frontier‑model AI ecosystem, clear and accessible reporting will only grow more important.

Further analysis of Frontier AI Frameworks and Transparency Reports from other frontier-model developers to follow in the next installments of this series. Stay tuned for what's next.

Readers with questions or corrections, please contact info@friartek.com.

Legal Disclaimer  

This blog post is provided for general informational purposes only and does not constitute legal advice, nor does it create an attorney‑client relationship. The analysis and summaries of the Transparency in Frontier Artificial Intelligence Act (TFAIA) are not exhaustive and may not reflect the most current legal developments. Readers should consult qualified legal counsel for advice regarding their specific circumstances or compliance obligations.