AI Policy

Axios reports that Treasury Secretary Scott Bessent joined a meeting between White House Chief of Staff Susie Wiles and Anthropic CEO Dario Amodei on Friday, as the company continues litigation against the Pentagon over a "supply chain risk" designation. Sources described next steps focused on non-Defense agency access to Anthropic's Mythos Preview model.

According to The Washington Post, Anthropic chief executive Dario Amodei met White House Chief of Staff Susie Wiles in a meeting framed around government concerns over Anthropic's Mythos system, which the Post reports can automate portions of the work involved in carrying out cyberattacks. The Post reports that the meeting took place amid what it describes as a government reaction to the capabilities of the Mythos model. The outlet does not report that any formal regulatory action, executive order, or security review has been announced as an outcome of the meeting.

Anthropic has previewed a new cybersecurity-focused model, code-named Mythos, and CEO Dario Amodei reportedly held a meeting at the White House on Friday, according to a report published April 17, 2026 by The Verge. The Verge frames the preview and the reported meeting as part of a shift in Anthropic's relationship with the Trump administration, a relationship the outlet character

A confrontation between the U.S. Department of Defense and AI company Anthropic — sparked by Anthropic's refusal to allow autonomous military targeting and domestic surveillance uses of its models — has escalated into a formal supply-chain-risk designation and a sweeping ban on federal contractors doing business with the company. Writing in IEEE Spectrum, policy analysts argue the episode reveals that binding democratic oversight of military AI, not executive ultimatums or corporate ethics policies, is urgently needed.

OpenAI has published a policy paper outlining its vision for US industrial policy in what it calls the "Intelligence Age," proposing government action on AI infrastructure, economic opportunity, and institutional resilience. The document, released on the OpenAI blog, frames AI development as a national-level economic project requiring coordinated public and private effort. It represents the company's most direct public pitch to policymakers on how the US should govern and invest in AI's expansion.

OpenAI has released a detailed public document called the Model Spec, outlining the behavioral principles governing its AI models. The framework attempts to balance safety constraints, user autonomy, and operator accountability, and serves as a rare instance of a leading AI company formalising — and publishing — the value hierarchy embedded in its systems. The document is advisory in nature, with enforcement dependent entirely on OpenAI's internal training and deployment processes.

A new paper published on arXiv argues that AI fabrication of case law and legal citations is not a random error but a deterministic, predictable failure mode rooted in the architecture of transformer-based models. The researchers contend that lawyers, courts, and regulators must update their understanding of how these systems fail — and that existing professional duty-of-competence standards already have direct legal implications for attorneys who rely on AI-generated legal research without adequate verification.

OpenAI has published a blog post outlining the safety measures built into Sora 2, its latest video generation model, and an accompanying social creation platform called the Sora app. The company says its approach is "anchored in concrete protections," though the post provides limited technical detail on what those protections entail or how they will be enforced.

OpenAI has published prompt-based teen safety policies designed for developers using its open-source safeguard model, gpt-oss-safeguard. The policies provide age-specific moderation guidance to help third-party builders reduce risks when their AI products are used by minors. The release represents OpenAI's attempt to extend child safety standards beyond its own consumer products into the broader developer ecosystem.