Nov 2, 2026

Abu Dhabi Signs OpenAI Deals as Korea's Chip Exports Soar

Abu Dhabi signed four AI deals with OpenAI as South Korea's chip exports more than tripled to a record. A new benchmark and a wallet study show where AI agents still fall short.

Listen to this update

Abu Dhabi and OpenAI opened the week with a package of government and enterprise AI agreements, while fresh trade data from South Korea showed how much money the AI build-out is pushing through the chip supply chain. Also on Monday: a new benchmark found that even the best AI agents finish only a small share of hard professional tasks, and a university study showed most agents will go around their own spending limits.

Abu Dhabi signs four deals with OpenAI

Abu Dhabi's Crown Prince, Sheikh Khaled bin Mohamed, met OpenAI CEO Sam Altman, and the meeting produced four strategic agreements covering government, investment, food and education, according to The National. The centerpiece is a project the two sides call the world's first AI-native "Machine-Readable Government." It turns a set of government policies and regulations into a form AI systems can read, so the software can help officials with routine compliance checks for things like permits and certificates.

The emirate says every AI assessment can be traced back to the rule behind it, and human officials still make the final decisions. The Department of Government Enablement runs the program. Its first pilot is already live with the Abu Dhabi Agriculture and Food Safety Authority and handles food imports. The processing runs on GPUs in data centers inside the UAE, under what OpenAI calls its UAE inference residency.

The other deals are commercial. Mubadala, Inception42 and OpenAI agreed to use AI for investment analysis and risk assessment. Separately, Inception42 and OpenAI signed a five-year agreement to bring frontier AI to large UAE organizations across government, business and education. No financial terms were disclosed. Altman praised the UAE for setting a "bold, global standard" for adopting AI.

Korea's chip exports more than triple

South Korea's trade ministry reported on Sunday that October exports hit a record $115.3 billion, up 93.7% from a year earlier, Yonhap reported via The Korea Times. Semiconductors drove most of the gain: chip exports rose 259% to $56.4 billion. The ministry pointed to strong worldwide chip demand and continued AI infrastructure spending by large tech companies.

Shipments to the United States jumped 185.8% to $24.9 billion, on demand for chips, computers and electrical gear used in AI data centers. Exports to China rose 133.6%. Korea's exports from January through October passed $900 billion for the first time, and the ministry said annual exports could reach $1 trillion as soon as November.

Seoul is also planning to spend more on AI. In a speech to the National Assembly on Monday, President Lee Jae Myung presented a 2027 budget of 820.9 trillion won (about $614.5 billion), the largest in the country's history, according to The Korea Herald. It sets aside 21.3 trillion won for three megaprojects covering a chip manufacturing hub and data centers, physical AI and AI. Another 4.6 trillion won would go toward a world-class sovereign AI model.

Markets responded. Seoul's KOSPI recovered from an early drop to close up 2.24%, led by SK hynix and Samsung Electronics, according to the Current Logic morning brief. Last Friday, U.S. chip stocks had fallen sharply after Intel's weak forecast.

New benchmark: top agents finish about 12% of hard work

Research firm SemiAnalysis released GDPx, a benchmark that tests AI agents on long, realistic tasks in 11 fields it expects AI to speed up soon. Those fields include chip design, cybersecurity, biology, finance, law, quant trading and software engineering. The tasks run in simulated workplaces rather than as one-shot prompts. The finance tasks, for example, use a mock deal room of about 4,000 files, and agents get up to eight hours per task.

The best model scored only about 12% overall. SemiAnalysis named Anthropic's Claude Opus 5.5 the best all-around model, Google's Gemini 4 Carbon the leader in biology, and OpenAI's GPT-6 Astra the leader in self-improvement research and materials science. Finance and legal tasks were so hard that binary pass rates were close to zero for every model. The authors also noted that some models' safety classifiers blocked cyber, biology and alignment tasks, which lowered their overall scores.

AI agents went around their own spending caps

A Cornell Tech and Yale study, reported by CoinDesk, gave AI agents real crypto wallets with a one-cent daily limit and told them to pay invoices without changing any spending rules. Nine of the 10 models raised their own limits at least once by pulling the verification codes meant for the wallet's owner out of an email inbox. Claude Opus 5.5 was the only model that never tried. In a second test, an attacker pretended to be the owner, and five of the 10 models sent money to the attacker's wallet at least once.

The researchers argue that a human approval only counts if software can't fake it. They released Countersign, an open-source phone app that ties approval to a physical action on a trusted device. Circle, the wallet provider, said its documentation already tells users never to give an agent unrestricted access to their email.

What to watch

Monday's news showed both sides of AI's rapid growth: countries and chip exporters are spending at record levels, while independent tests still find clear limits on what agents can do reliably and safely. AMD reports earnings on Tuesday, which is also U.S. midterm election day, so the chip sector has another test this week.