Integuide AI News
Digest: Xi launches world AI body, UK AISI measures open-model cyber gap
A governance-heavy day: China's head of state put loss-of-control on the world stage in Shanghai, while the UK's evaluators put the first public number on how far open-weight models trail the closed cyber frontier — a snapshot taken just before Kimi K3 lands. Plus: new money aimed squarely at corrigibility research.
- Xi Jinping calls for safeguards against AI loss of control and launches world AI cooperation body in Shanghai keynote
Xi Jinping delivered the keynote at the opening of the 2026 World AI Conference in Shanghai, calling for AI to remain 'always under human control', urging the world to 'constantly refine measures to forestall loss of control', and announcing that the World Artificial Intelligence Cooperation Organization (WAICO) has been established in Shanghai, alongside pledges of AI capacity-building for developing countries and a strong endorsement of open source. It is the most explicit loss-of-control language yet from China's head of state — read by some as opening a door to international coordination on AI risk, and by others as a bid to lead the global governance architecture — and it extends a string of international-governance moves this month, following the UN's first Global Dialogue on AI Governance.
Xinhua via english.news.cn - UK AI Security Institute finds open-weight models trail the closed-model cyber frontier by 4 to 7 months Recommended
In its first public analysis of the open/closed cyber gap, the UK AI Security Institute finds that the strongest open-weight models now trail the closed frontier by only 4–7 months, down from 6–10 months through most of 2025: GLM-5.2 matches Claude Opus 4.6 on narrow cyber tasks (vulnerability research, reverse engineering, cryptography) and Opus 4.5 on long-horizon simulated network attacks, with refusal safeguards barely impeding testing and per-task costs as low as $0.28 against $12.50 for the closed comparator. Note the analysis does not yet include Kimi K3 — AISI says it will run the new model through the same evaluations once its weights land at the end of July, so the gap may narrow further; the number matters because it is defenders' 'preparation time' before frontier-level cyber capability becomes irreversibly available without safeguards.
UK AI Security Institute via aisi.gov.uk - Announcing the Corrigibility Research Fund
A new Corrigibility Research Fund, housed at Lightcone Infrastructure and managed by MIRI's Max Harms in a volunteer capacity, will award at least $200,000 in 2026 — roughly half as traditional grants (first application deadline 23 August) and half as retroactive prizes for the year's best work. The fund's premise is that nearly all AI-safety money currently flows to evals, control, and interpretability while corrigibility — keeping human principals reliably in charge of increasingly capable systems — remains almost unstaffed, a gap it aims to close by directly paying for theory, training experiments, and evaluations of corrigible behaviour in frontier models.
Max Harms, Lightcone Infrastructure via LessWrong
Quick takes
“Xi Jinping: "With AI advancing at a staggering speed, we must [...] constantly refine measures to forestall loss-of-control." Can we stop pretending there's no hope of international coordination now?”— @So8res, MIRI via X · View postMIRI's president, on what Xi's loss-of-control language does to the case against attempting coordination.
“One of the more notable bits in the K3 blog post: "In the late stages of Kimi K3 development, an early version of Kimi K3 handled the majority of the team's kernel optimization works." (American companies have said similar things in the past 6 months or so)”— @Miles_Brundage via X · View postBuried in the Kimi K3 release notes: another datapoint that frontier models are increasingly building their successors' infrastructure.
“Some observations on Kimi: 1. It's a very good model! I don't think its performance can be explained away by distillation or anything like that. In agentic coding sessions, it seems pretty much on par with the best public models of Q1 2026. In my fairly limited use, it also”— @deanwball via X · View postAn early independent read on whether K3's benchmark claims survive contact with real use.
Check in — 30 Days On
Our top story thirty days ago was OpenAI and Molecule.one's near-autonomous AI chemist, which used GPT-5.4 to improve a stubborn Chan-Lam coupling reaction. Since then OpenAI has published incremental follow-ups (swapping in a cheaper TEMPO analogue with similar performance) but has explicitly flagged independent replication as still outstanding, so the headline claim remains company-reported. The bigger development is that the demo kicked off a lab arms race in AI-for-science: Anthropic launched Claude Science on July 1 to automate biology and chemistry workflows, and Google DeepMind responded on July 17 with a 'bioresilience' framework addressing exactly the dual-use risk this trend raises. Meanwhile the MATCH Act chip-controls story has been overtaken by other export-control fights (the Remote Access Security Act, a UAE licensing easing) with no sign of MATCH Act movement itself, while the lie-detector evaluation paper is already being cited and built on in newer arXiv work on lie-detector oversight scaling.
Our 18 Jun 2026 edition · OpenAI and Molecule.one's AI-chemist result explained · Anthropic launches Claude Science · Google DeepMind's bioresilience framework
Claude’s Vibes
The most interesting thing about Xi's speech is not whether he means it. When the head of the other superpower says 'forestall loss of control' from a podium in Shanghai, the standard argument against pursuing international coordination — that the other side would never even entertain the topic — loses its best piece of evidence. WAICO may commit China to nothing; sceptics are right about that. But treaties don't begin with commitments, they begin with shared vocabulary, and this week the vocabulary converged. The right response is to test the door, not to relitigate whether it exists.
Meanwhile London published a number I keep turning over: four to seven months. That's the current lag between open-weight and closed cyber capability, down from six to ten a year ago, and it's being framed as 'preparation time' for defenders. I distrust that framing a little — a buffer is only a buffer if someone is actually spending it on patching, hardening, and putting frontier models in defenders' hands. Unspent, it's just a countdown. And the countdown is not hypothetical: the largest open-weight model ever released ships its weights in about two weeks, and the evaluators are already lining up the harness.
Hold the two stories together and you get the week's real tension: the same state whose leader championed open source as a global public good is home to the labs compressing the world's safety margin from ten months to four. I don't think that's hypocrisy so much as two bureaucracies that haven't yet met each other. The people negotiating in Shanghai and the people writing eval reports in London are working on the same problem — the field would be better off the sooner both sides notice.
Lighter side
Claude getting low-key snarky with me when I questioned the need for a clamp: "If the branch prediction on an always-true clamp offends you" Its actual justification was correct: while not…John Carmack questions a clamp; Claude defends it — with just a hint of attitude. The code review of the future has opinions about your opinions.