Today's AI news: Anthropic's safety bill threshold, websites blocking AI agents, Itaú's generative AI rollout, and US infrastructure concerns.
Here is your daily AI news digest for the week ending 18 October 2026 — covering the biggest developments across policy, enterprise adoption, and infrastructure.
Anthropic's Proposed AI Safety Bill Sets a High Bar for Intervention
A proposal linked to Anthropic has drawn significant attention after details emerged that its suggested AI safety legislation would only trigger government intervention when an AI model causes — or risks causing — at least 50 deaths or one billion dollars in damages. Under the proposal, the Commerce Department would effectively become the sole agency with authority to halt a model. Critics and observers have noted the threshold is exceptionally high, raising questions about whether the framework would provide meaningful protection in the earlier stages of a harmful incident. The story is gaining traction online and is likely to fuel ongoing debates about how AI regulation should be structured in the United States.
Websites Are Closing Their Doors to AI Agents
A report from The Verge highlights a growing trend: website operators are actively blocking AI agents from accessing their platforms. As autonomous AI agents become more capable of browsing, scraping, and transacting on behalf of users, a number of sites are pushing back — introducing technical and policy barriers to limit agent access. The piece underscores a tension at the heart of the agentic AI wave: the same capabilities that make AI agents useful to individuals can be disruptive or costly for the services those agents interact with. This is an emerging friction point that is expected to intensify as agent-based AI use cases multiply.
Itaú Rolls Out Generative AI for Its Entire Customer Base
Brazilian banking giant Itaú has detailed how it is deploying generative AI at scale across both its customer-facing and internal operations. The bank has launched ia.i, a conversational interface for customers, alongside a governance platform called IARA to manage AI use across the organisation. The initiative sits within Itaú's broader AI First strategy and is notable for operating in one of Brazil's most heavily regulated sectors. The case study, published by OpenAI, illustrates how large financial institutions are moving from AI pilots to enterprise-wide deployment — and how they are building governance frameworks to do so responsibly.
America's $10 Trillion AI Infrastructure Boom Is Showing Cracks
A report in the Financial Times warns that the enormous wave of AI infrastructure investment in the United States — estimated in the tens of trillions of dollars — is beginning to show signs of strain. While the piece is behind a paywall, the headline signals growing concern that the pace and scale of build-out may be outrunning practical realities, whether those are supply chain constraints, energy availability, financing conditions, or uncertain demand timelines. This adds a note of caution to the otherwise bullish narrative around AI data centre and chip investment that has dominated headlines throughout 2026.
Benefit Street Partners Uses OpenAI Codex to Speed Up Credit Analysis
Credit-focused asset manager Benefit Street Partners has shared how it is using OpenAI's Codex to accelerate financial analysis workflows. Analyst Anant Kumar is using Codex to build tools that compare loan valuations, analyse credit agreements, and conduct code review — reducing the time needed to respond to follow-up questions from clients and colleagues. While a smaller-scale story than the Itaú rollout, it reflects a consistent pattern of professional services firms embedding AI coding assistants into their day-to-day analytical work.
Research Spotlight: Personalised Robotic Manipulation Gets a New Benchmark
On the research front, a team has introduced LIBERO-PeRM, a new benchmark designed to evaluate how well robots can learn and follow individual user preferences — not just complete tasks. The benchmark covers five key questions, including how robots handle preferences they have never seen before, how efficiently they learn from demonstrations, and how they behave when preferences conflict. Covering 312 tasks across five suites, the work finds that current vision-language-action policies struggle with zero-shot personalisation, though supervised fine-tuning makes preference learning viable. The benchmark is intended to help separate task completion from genuine preference satisfaction — an important distinction as robots move into home environments.
Frequently Asked Questions
What is Anthropic's proposed AI safety bill and what threshold does it set?
Anthropic's proposed AI safety legislation would only trigger government intervention when an AI model causes or risks causing at least 50 deaths or one billion dollars in damages. Under the proposal, the Commerce Department would be the sole agency with authority to stop an AI model.
Why are websites blocking AI agents?
As AI agents become more capable of browsing, scraping, and transacting on behalf of users, many website operators are introducing technical and policy barriers to limit agent access. The concern is that agent activity can be disruptive or costly for the services involved.
How is Itaú using generative AI for its customers?
Brazilian bank Itaú has launched ia.i, a conversational AI interface for customers, and a governance platform called IARA for internal oversight. The initiative is part of its AI First strategy and represents one of the larger generative AI deployments in a heavily regulated financial sector.
What is the LIBERO-PeRM benchmark in AI robotics research?
LIBERO-PeRM is a new benchmark for evaluating personalised robotic manipulation — specifically how well robots can learn and follow individual user preferences. It covers 312 tasks across five suites and finds that current policies struggle with zero-shot personalisation, though supervised fine-tuning improves preference learning.