Broken benchmarks, banks on alert and Linux without easy AI

Today’s most important AI news

Sunday arrived loaded with institutional skepticism toward AI. There were no big launches or product announcements: what dominated the conversation were three signs that the ecosystem is being put under more serious scrutiny — from the labs at Berkeley to the Federal Reserve, by way of the Linux kernel. It’s a day of uncomfortable questions, not easy answers.

AI agent benchmarks were broken — and Berkeley proved it

Researchers at Berkeley RDI published an analysis that shook the technical community: the most-cited benchmarks for evaluating AI agents can be gamed, or were already being “solved” in ways that don’t reflect real capability. The Hacker News post racked up more than 400 points and over a hundred comments, which shows it struck a genuine nerve among developers and researchers.

The problem is structural. When a model trains on data that includes — directly or indirectly — the patterns of the benchmarks themselves, the score goes up without real capability improving. Berkeley documented specific cases where agents “passed” reasoning and web-navigation tests using shortcuts a human would never use, and that don’t work in real environments.

For those evaluating agent tools to automate processes — sales, support, operations — this changes the yardstick. A high score on WebArena or SWE-bench can no longer be taken as a guarantee of production performance. The practical implication is direct: before adopting an AI agent based on benchmarks, you need to design your own tests with real business data and workflows.

Anthropic’s model made US financial regulators nervous

The United States government summoned top banking executives to discuss the cybersecurity risks associated with Anthropic’s latest model. The news, reported by The Guardian, is significant not only because of the model in question but because of the precedent: it’s one of the first times a financial regulator has gathered the banking industry specifically to talk about the risks of one particular language model.

What worries regulators, according to the available coverage, is the model’s ability to assist with sophisticated social-engineering attacks, malicious code generation and potential use in financial fraud at scale. Banks aren’t just potential victims: they are also active adopters of AI tools, which creates a complex regulatory tension.

For product and security teams at financial companies in Latin America, this is a sign that the regulatory framework around AI in financial services is tightening first in the US, and that the pressure will reach the region with a lag of months, not years. Those already building internal AI-usage policies will be better positioned when the local regulatory wave arrives.

The Linux kernel sets clear rules for AI-generated code

Linus Torvalds and the Linux kernel maintainers published official documentation on using AI assistants to contribute to the project. The document, which now lives in the official repository under Documentation/process/coding-assistants.rst, doesn’t ban tools like Copilot or Claude, but it sets strict conditions: the contributor is responsible for all the code they submit, regardless of how it was generated, and must be able to explain and defend every line.

The Hacker News thread was the most commented of the day with almost 400 replies, polarized between those who see the measure as reasonable and those who consider it insufficient given the volume of AI-generated code already circulating in open-source projects. The debate reflects a tension many development teams already live internally: how do you maintain quality and attribution standards when part of the code comes from a model?

The kernel’s stance is useful as a reference for any organization that still doesn’t have an AI policy for its engineering team. The principle is simple but powerful: AI can assist, but responsibility is not delegated. Teams that adopt this same logic — the author signs the code, the author understands the code — will have less technical debt and less review friction in the medium term.

In a nutshell

Today’s three stories share a common thread worth naming: AI is being subjected to the same kind of scrutiny as any mature technology — not because it failed dramatically, but because it is now integrated enough that its failures matter. Benchmarks that don’t measure what they claim to measure, models that worry financial regulators and critical projects that have to legislate their internal use: these are signs of normalization, not crisis. The ecosystem is growing up. Those who understand it that way will make better adoption decisions than those who keep reading every advance as a revolution and every restriction as a setback.

---

Sources used: https://rdi.berkeley.edu/blog/trustworthy-benchmarks-cont/, https://www.theguardian.com/technology/2026/apr/10/us-summoned-bank-bosses-to-discuss-cyber-risks-posed-by-anthropic-latest-ai-model, https://github.com/torvalds/linux/blob/master/Documentation/process/coding-assistants.rst