Tomorrow, June 25, UNESCO’s Third Ministerial Summit on the Ethics of Artificial Intelligence in Latin America and the Caribbean opens in Santo Domingo, organized together with the Dominican Republic’s OGTIC and the development bank CAF (OGTIC). It is the first regional ministerial forum after a month in which the major AI governance decisions were made without the region in the room: the US ban on the Fable 5 and Mythos 5 models (June 12), a G7 in Évian that closed without a common AI framework (June 17) and a Colombian election decided by less than one point with active digital interference on election day (June 21).
The ministers of technology and digitalization arrive with two unprecedented facts on the table that were not on the original agenda, and with just 48 hours to produce a declaration, scheduled for June 26. The underlying question is whether that document will name what the region lived through this month —or whether it will repeat the principles that have been circulating for three years without a mechanism to enforce them. The summit cannot reverse any of those three facts, but it could set what minimum standard the region would commit to having operational before Brazil’s elections in October 2026, and what consultation channel it would demand in the face of the next “off switch” decided in Washington (UNESCO).
The technical backdrop arrived at almost the same time. Google DeepMind published a roadmap for controlling AI agents with fifteen “defense in depth” controls, built on a principle that clashes with the usual discourse of accelerated adoption: AI agents themselves should be treated as an insider threat, with the same safeguards that would apply to a high-privilege outside contractor. The document sets, in practice, what “deploying AI responsibly” means in 2026 —a standard considerably more demanding than any public framework in force in the region. If a Latin American public institution wanted to use frontier models in an audited and sovereign way, that is the architecture it would need to understand; the uncomfortable question is how many of those fifteen controls it could implement in the next twelve months.
Also today
- Google DeepMind publishes its roadmap for controlling AI agents — Fifteen defense-in-depth controls that treat models as insider threats; it defines the technical standard that appears to be under negotiation as a condition for lifting the ban on Fable 5.
- Gemini 3.5 Pro did not debut on the first day of its launch window — The confirmed window (June 23 to 30) passed its first day without general availability; it remains in limited preview. Prediction markets dropped to 50-55% for before June 30.
- OpenAI expands its Daybreak cybersecurity initiative with IBM and Tenable — It adds two top-tier enterprise partners (IBM, with an active presence in the region) and launches “Patch the Planet,” an open project to fix vulnerabilities automatically.
- Stack Overflow opens its 2026 survey “for human developers only” — The explicit framing acknowledges that the previous edition was contaminated by AI-generated responses; a signal about the very epistemology of industry surveys.
In the region
Santo Domingo is the first test of whether Latin American multilateralism can produce concrete instruments and not just declarations of principles. The structural risk emerging this week is new: if the way out of the Fable 5 ban requires verifying US residency for a first tier of access, access to frontier models differentiated by citizenship could become the norm rather than the exception. It would be the first precedent of an AI export control applied not to chips or hardware, but to inference itself, with discrimination by nationality built in. For the ministers, that is the most important governance question of the week: not whether the United States can control its exports, but whether the region has any mechanism to demand access conditions equivalent to those of a US citizen.
Threads we’re following
The US ban on Fable 5 and Mythos 5 —which we have been following since it was announced on June 12— reaches its twelfth day with the most visible exit architecture so far. An executive order that would exempt Anthropic from export controls, conditioned on residency verification, converges with an update to its privacy policy (in effect from July 7) that enables biometric citizenship verification. Together they shape a two-speed restoration: the United States first, everyone else later. Prediction markets give a 57% probability of resolution before July 1. If the mechanism requires US federal documentation, Latin American users would not be in the first tier —and that scenario is still open as the ministers sit down tomorrow in Santo Domingo (status of the ban).
If the June 26 declaration could contain a single concrete instrument —implementable and verifiable before October 2026— that responded both to access to frontier models without regional consultation and to the detection of disinformation without sanctions, what would it be, who would carry it out and what would make it binding?
About this entry. It is generated automatically from public sources, without human review before publication. It may contain errors of interpretation or summary; please check each story against its original source (the links lead there) before citing it or making decisions based on it.
Doble Click is written with Anthropic models.