Overview

The strongest signals sit at three different layers of the AI stack. Anthropic is contracting directly for memory, Daybreak is pairing model discovery with expert remediation, and CORE-Bench shows that saturated accuracy can conceal shortcuts, inefficiency, and unreliable scaffolds.

1. Micron and Anthropic sign an AI infrastructure supply agreement

Micron and Anthropic signed an AI infrastructure supply agreement, Reuters reported. The deal connects a frontier-model developer directly with a memory supplier, reflecting how high-bandwidth memory and related components have become strategic inputs instead of interchangeable purchases made only by cloud operators.

The disclosed agreement omits volumes, duration, pricing, and delivery regions, preventing a capacity or unit-cost estimate. Direct contracting nevertheless moves a frontier laboratory closer to a component supplier and makes memory availability part of Anthropic's model-serving plan instead of solely its cloud providers' procurement problem.

Sources: Reuters

2. Daybreak pairs model-generated findings with expert validation and patches

OpenAI's Patch the Planet program embeds Trail of Bits engineers across 19 open-source projects to validate AI-generated findings, prepare patches, and coordinate disclosure. The initial work had identified hundreds of issues and merged dozens of patches when OpenAI published its report.

Reported outputs include eight Linux kernel information-leak proofs of concept, 24 local privilege-escalation exploits, and four patterns corresponding to six later-fixed dnsmasq vulnerabilities. Every report reached a security engineer before a maintainer, so validated fixes, not raw model alerts, define the program's output.

Sources: OpenAI's Patch the Planet report · Trail of Bits announcement

3. CORE-Bench remains informative after accuracy saturates

Researchers revisited CORE-Bench Hard after AI agents approached its accuracy ceiling and found threats to construct validity that weaker systems had not exposed. They released a corrected v1.1 suite and an out-of-distribution set for computational reproducibility tasks.

The study evaluates six dimensions beyond accuracy: shortcuts, out-of-distribution transfer, efficiency, reliability, model contribution, and scaffold contribution. A small randomized experiment also found human-agent teams completed reproducibility work about twice as fast as humans alone, with one-fifth of human-only attempts timing out.

Sources: CORE-Bench analysis