#2 - We Hope And We Go To Work

By Ima, José and Yannick

As AI capabilities continue to advance rapidly on both sides of the Pacific Anthropic’s Mythos made headlines due to its threat to global cybersecurity. Is this surprising? Not really. Mythos is built on a constant trend of improving performance that follows naturally from previous models. But does that mean it’s not concerning?

No. Mythos demonstrates measurable progress in autonomous cybersecurity tasks, including vulnerability discovery and multi-stage attack execution in controlled settings.

This means defensive preparations, notably for critical infrastructures, need to improve, and quickly. Otherwise, systems across the world - from bank accounts to power grids, water treatment plants to telecommunication networks - risk being compromised.

Keeping Up With Global AI Governance

Governments

At Davos, Mark Carney’s speech highlighted a “rupture” in the rules-based world order, blaming unconstrained great-power geopolitics. He also urged middle powers (such as Canada, and EU nations) to unite for strategic autonomy, human rights and resilience. The speech resonated amid tech sovereignty debates.

At the Delhi Summit, India asserted the Global South voice in AI governance. Three things to point out:

(1) Scale: The Summit witnessed extensive participation, from more than 100 countries.

(2) Framing: Just like France, India approached its summit as a trade event, with– as Amber Sinha put it–”one clear economic objective outweighing all others — to seek private sector investment into domestic AI projects.”

(3) A new platform: The Summit launched the civil-society led Global South Network for Trustworthy AI, tasked with making the New Delhi Frontier AI Impact Commitments operational. The two commitments focus on evidence-gathering around AI’s economic impact and on evaluation capacity for under-represented languages and contexts Despite the best efforts of civil society, this might not be sufficient to influence advanced models’ training, deployment and procurement, but it definitely is a step in the right direction.

Perhaps the most iconic image of the entire summit was the sight of two men (Sam Altman and Dario Amodei) refusing to hold hands, while presenting these commitments to the world.

The next summits will be hosted by Switzerland and the UAE in 2027 and 2028 respectively.

International Organizations

Early March, the UN General Assembly appointed 40 independent experts to the Independent International Scientific Panel on AI (ISP-AI) with Yoshua Bengio and Maria Ressa elected as co-chairs at the group’s inaugural meeting. Designed as an “IPCC for AI”, the Panel will produce annual scientific reports on the risks, opportunities and developmental dimensions of AI development, without regulatory authority. As the members disperse after Madrid, we wish them well in the hard work of balancing perspectives into a report that reflects concerns and hopes worldwide.

They won’t be working in a vacuum. Even before Mythos or the Anthropic vs US Department of War dispute, a growing number of states (Spain, Brazil and other BRICS nations, Bahrain, Cote d’Ivoire and Saint Vincent and the Grenadines) had made clear that a robust global governance framework was needed. A global scientific report feels like a stepping stone towards one.

Ahead of the first UN Global Dialogue on AI Governance, we will publish a short series explaining our building blocks for international AI governance.

Stay tuned!

What’s Up With Those Chips?

Compute is where AI governance is less abstract. Who makes the chips, who gets them, and whether they can be turned off are now live policy questions, and a rare area where the US, China and middle powers share real interests.

Start with the off-switch. President Trump recently floated the idea of an “off-switch” for advanced AI. The policy value is straightforward: a credible, verifiable way to pause or constrain specific deployments is one of the few tools that works across rival jurisdictions without requiring them to trust each other.

Two other developments matter for policymakers:

• G42’s Assurance Compute Framework (Feb 2026) — the first serious attempt by a major non-US buyer to operationalize the State Department’s Pax Silica initiative.

• Quantum’s “inflection point” according to some people. Hardware-based AI governance tools rely on cryptography. If post-quantum transitions lag, the verification infrastructure underpinning export controls and compute accountability weakens with it. So, to avoid existing cryptographic systems being broken, the industry needs to transition to so-called post-quantum cryptography.

The Oxford Martin Hardware AI Governance Lab (Awad & Trager, with our own James and Yannick as Research Affiliates) published its first report on trusted semiconductor manufacturing. This is a good entry point if you want to go deeper.

Keep Your Eyes Peeled

🇸🇬 International Scientific Exchange on AI Safety, Singapore, 17–19 May. Technical researchers, AISIs, frontier labs. Focus: Global Safety Research Priorities and agentic risk management. Why it matters: one of very few venues that brings together key technical researchers from across the world, including representatives from AI safety institutes and the leading AI companies.

🇨🇳🇺🇸 Trump–Xi, Beijing, May (postponed from earlier). First of several meetings this year. AI guardrails expected on the agenda. Why it matters: Cooperation on guardrails around AI development is expected to figure in the agenda.

🇺🇳 UN Global Dialogue on AI Governance, Geneva, 6–7 July. Co-facilitated by El Salvador and Estonia. Four thematic clusters: opportunities, divides, safety, human rights. Why it matters: first operational test of the new UN architecture.

🔬 ISP-AI first interim report, due ahead of the Global Dialogue. Chaired by Yoshua Bengio and Maria Ressa. Why it matters: the ISR Report, despite scientific legitimacy backed by political will, lacked global momentum.

We’ll be back next month. Don’t forget to visit global-governance.ai for an analysis of current efforts, gaps and recommended reading.

In the meantime, let’s hope, and go to work.