AI This Week: Anthropic Ships Claude Sonnet 5, Japan Bets ¥1 Trillion on Robot Workers, Bank of England Wonders If Finance Rules Fit

Anthropic released Claude Sonnet 5 and turned Fable and Mythos back on after an eighteen-day pause. The US federal export control review came after Amazon researchers found a way around Fable 5 safety controls. Anthropic locked down the most powerful systems for all users while they worked on a fix.

Anthropic built a safety classifier for the specific bypass method. It now catches over 99 percent of attempts. When a prompt hits the safety boundary, the system sends the work to the older Opus 4.8. The incident shows the tension between powerful AI and government oversight. Security teams found similar weak spots in GPT-5.5 and Kimi K2.7.

Claude Science and NVIDIA BioNeMo

Anthropic's Claude Science workbench now works with the NVIDIA BioNeMo Agent Toolkit. Scientists describe their research in plain language. Genome analysis, protein structure prediction, and molecular binder design happen through Claude Science with domain-specific agents on NVIDIA's GPU stack. Eighteen of the top twenty drug companies already run BioNeMo in production. This puts AI at the center of computational biology.

Japan's ¥1 Trillion Robot Plan

Japan moved from talk to action on robot AI. The government hired Noetra, a group owned by SoftBank, NEC, Sony Group, and Honda, plus the national lab AIST to build a "physical AI" model. It reads language, images, video, and sensor data together. The target is 10 million AI robots across 18 industries by 2040, with up to ¥1 trillion (about US$6.1 billion) in public money over five years.

The project runs from fiscal 2026 to 2030. Only the first two years have guaranteed funding. Tokyo can pull the plug if targets slip. An early version of the model could come this fiscal year.

Bank of England Looks at AI Agents in Finance

Deputy Governor Sarah Breeden told the European Central Bank Forum that current rules were "not built to contemplate autonomous agents in payments, trading, and operational functions." A Cambridge Centre for Alternative Finance report found 52 percent of financial firms already use AI agents. The Bank of England wants to know if the rulebook covers systems that make decisions and act without a person in the loop.

Breeden pointed to cyber resilience. Agentic systems can chain actions faster and at bigger scale than older automated tools.

HP Rolls OpenAI Frontier Across the Company

HP put OpenAI Frontier into its global operations after pilots showed big productivity gains. One engineer handled 122 pull requests across 43 projects in weeks. The security team fixed bugs in a single day that would have taken a month before. HP uses ChatGPT for broad knowledge tasks and keeps code-generation models for engineering work. Other companies can use this as a template for large-scale AI adoption.

What It All Means

Three threads came together this week. AI moves from experiments into real operations in pharma, finance, manufacturing, and national infrastructure. Governments respond. The Fable 5 export review, the Bank of England rule check, and Japan's milestone-based funding all show that regulators are active participants now. For anyone in business, the time to build an AI strategy and the governance around it grows short.