AI WORLD SIGNAL

ARCHIVE / EN

AI intelligence archive

Source-linked analysis connecting models, infrastructure, companies and markets.
Latest intelligence2026-07-28

Safety · Agent security · Research

GPT‑Red attacks AI agents: why a 0.05% failure rate does not mean prompt injection is solved

OpenAI trained an automated attacker with reinforcement learning and self-play, then used its attacks to harden GPT‑5.6. The results are strong, but largely vendor-run, with thinner coverage of multimodal and long-horizon attacks.Read the analysis
Previous intelligence2026-07-27

Infrastructure · Korea · Sovereign capital

From 55MW to 200MW: what is firm—and what is conditional—in Korea’s $10B AI factory plan

NAVER, NVIDIA and Brookfield are combining a data center, accelerated compute, private capital and open-model work in one plan. It is not a completed build: part of the financing remains conditional, and Brookfield’s term sheet is described as nonbinding in the global release.Read the analysis
From the archive2026-07-26

Models · Agents · Compute economics

Claude Opus 5 resets the intelligence-cost curve: what Anthropic’s launch means for enterprises

Anthropic launched the model on July 24, 2026 with Opus 4.8 pricing, a one-million-token context window and adaptive thinking on by default. The important shift is not another benchmark lead; it is competition around completed-task economics and agent reliability.Read the analysis