AI Regulation & Policy

Today's Stories

1. 1

   U.S. lifts AI export controls on Anthropic's Fable model

   The U.S. government reversed export controls it had imposed two weeks earlier on Anthropic's Mythos and Fable models. The government first lifted controls on Mythos on Friday evening, then on Fable late Tuesday. Both models had been disabled for all users during the restriction period. The temporary controls exposed a risk that U.S. businesses now recognize — they cannot reliably depend on American frontier AI models for essential tasks without fallback options. This concern is pushing more enterprises to explore open source alternatives, even though the most capable open source models currently come from Chinese AI companies, creating a dilemma over reputational and geopolitical risks.

   The U.S. is working with leading AI labs on explicit voluntary cybersecurity standards that frontier AI models can meet to avoid government objection to public release. Anthropic announced it is also developing a shared framework with Amazon, Microsoft, Google, and other critical infrastructure partners to assess risks from jailbreaks to model guardrails.

2. 2

   Warp raises $60M to automate payroll and HR with AI

   Employee management startup Warp has raised $60 million（約96億円） in new funding. The New York-based company is positioning itself as an AI-native alternative to legacy human capital management software, using artificial intelligence to handle payroll and back-office work with minimal staff involvement. The payroll and HR software market has long been dominated by established players like Workday. Warp's approach suggests a shift toward AI-driven automation in administrative functions—a move that could reshape how businesses manage compliance and personnel costs.

   Warp pitches itself directly against the legacy systems that have controlled this category for years. The company's focus on reducing staffing needs for back-office operations will be key to proving whether its AI-native model can win over enterprise customers.

3. 3

   Enterprise AI lacks single owner, creating control gap as spending races ahead

   VentureBeat Pulse Research found that most organizations run multiple competing AI platforms, few can detect when a model fails in production, and the most-cited barrier to control is the absence of a single owner accountable for AI across the entire stack. Autonomous agents are already producing real financial and operational failures. AI portfolios are expanding faster than the ability to govern them. This widening control gap — where ambition and spending race ahead of visibility, ownership, and cost control — means enterprises are losing track of their AI investments and cannot reliably manage their behavior at scale.

   The research specifically examines how many platforms claim to be the primary AI layer, who actually governs AI behavior across them, what blocks cross-platform governance, and how financial and operational failures of autonomous agents are already surfacing in organizations.

4. 4

   Inscribe detects document fraud in 90 seconds using AI on AWS

   Inscribe built an AI system using Amazon Bedrock that detects tampered, fabricated, and AI-generated financial documents in under 90 seconds—a 20x improvement over traditional manual review, which takes 30 minutes per application. Financial institutions processing thousands of loan and credit applications daily face fraud in 1 of every 16 documents, with AI-generated forgeries growing 5x from April to December 2025. Manual review cannot keep pace with volume or detect sophisticated deepfakes and coordinated fraud rings, leaving institutions exposed to millions in losses and regulatory risk.

   Inscribe uses multiple models from Amazon Bedrock matched to specific tasks—Claude Haiku 4.5 for routine document parsing (40% cost reduction vs. Claude Sonnet), Meta Llama models for transaction analysis, and Claude Sonnet 4 and 4.5 for cross-document fraud pattern detection and audit-ready reporting.

5. 5

   AI safety experts favor treaty over 'CERN for AI' model

   A researcher argues that while a "CERN for AI" (a shared international research lab) sounds appealing, the politically realistic versions would not meaningfully improve AI safety, while versions that would help are probably unachievable. Instead, they propose an international treaty with enforcement mechanisms, modeled on how the EU AI Act, the Nuclear Non-Proliferation Treaty, and the Montreal Protocol actually developed. The core bottleneck in AI safety is not more research but political will and enforcement of best practices. A treaty-first approach, followed by an IAEA-style (International Atomic Energy Agency–style) verification body, could reduce safety risk substantially—roughly 80% if enough political commitment exists. This reframes the debate away from funding new labs and toward binding agreements between nations.

   The argument rests on the premise that enforcement and political will are the limiting factor in AI safety, not scientific breakthroughs. The proposal draws a parallel to how nuclear safety and environmental protection treaties have been structured historically, suggesting a sequencing that treats verification as the critical step.

6. 6

   Genpact launches AI tool to recover lost revenue for consumer goods firms

   Genpact has launched an AI-powered Deductions Recovery solution for consumer goods companies, using Microsoft Azure and specialized AI agents to automate deduction management and recovery processes. The product targets a concrete pain point—preventable trade deductions and unresolved invalid claims that create gaps in cash collection for consumer goods companies. By automating data aggregation and resolution, Genpact aims to expand its higher-margin, AI-centric transformation services and deepen relationships with existing clients' finance and operations teams.

   The speed at which the platform gains client adoption and whether it can scale fast enough to meaningfully offset slowing legacy business-process outsourcing services. Genpact's stock has declined 40.1% year to date and 37.7% over the past year, making this launch a test of management's AI strategy.