The AI Brief

Published June 18, the GDM AI Control Roadmap (v0.1) is a technical framework and companion policymaker guide describing how Google is securing the AI agents it deploys inside its own research organization. The core stance: alignment training is a primary defense but not a sufficient one. The roadmap adds a system-level layer that treats certain internal agents as untrusted by default, drawing on zero-trust cybersecurity principles rather than assuming behavioral compliance after training.

The framework maps threats across three buckets: loss of control (rogue internal deployments), sabotage of safety research, and direct harm such as data deletion or theft. Controls span 15 mechanisms, ranging from chain-of-thought monitoring and prompt injection resistance at the detection layer, up to real-time blocking and human escalation for the most dangerous potential actions. DeepMind built and tested an asynchronous monitoring prototype against one million internal coding-agent tasks; that run found most flagged issues stemmed from overzealous behavior rather than adversarial intent, though the company treats that as evidence of the need for monitoring, not evidence against it.

The roadmap is published explicitly as a potential industry blueprint, not just an internal specification. A companion policymaker document calls on AISIs (AI Security Institutes) and governments to facilitate pre-standardization work and avoid a fragmented landscape of overlapping frameworks. DeepMind warns the window for establishing global safety standards for agent systems is closing as agent deployment scales. Fortune's coverage notes this comes as automated systems already account for 57.4% of global web requests.

On June 18, OpenAI announced a health intelligence upgrade to ChatGPT powered by GPT-5.5 Instant, the same model update underpinning recent health feature expansions. The factuality figure is sourced from OpenAI's Global Physician Network, a 262-doctor panel assembled to evaluate health-response quality. The 71% drop in factuality issues is an internally measured, vendor-reported figure; no third-party peer review or independent clinical audit of the metric has been published. OpenAI's disclosure that ChatGPT fields over 230 million weekly health interactions is also self-reported.

The rollout is available to all free ChatGPT users, not only Plus or Enterprise subscribers, which is the structural decision worth noting. Prior health feature launches concentrated physician-evaluated improvements in paid tiers. Distributing the upgrade across all access levels maximizes the network-effect argument for ChatGPT as a health information layer while raising the regulatory question of what obligation, if any, attaches to AI health guidance reaching a billion-scale user base.

Announced June 19, the feature covers Claude chat, Claude Code, and Cowork on Team and Enterprise plans. Seven MCP providers support Enterprise Managed Auth at launch: Asana, Atlassian, Canva, Figma, Granola, Linear, and Supabase. Slack support is listed as coming soon. Early adopters named in the launch include Ramp, Webflow, and HubSpot. Okta also issued its own announcement confirming it is the featured identity provider, noting the protocol is built on Cross App Access (XAA), an open extension Okta leads that is now officially recognized within the MCP authorization specification under the name "Enterprise Managed Auth."

The governance implication is substantive. Prior to this, an admin could enable a connector organization-wide, but each user still had to authorize it in a separate step. Security teams could not centrally audit which users had authorized what. The new model pushes that control surface into the IdP, where enterprises already manage identity risk posture, access reviews, and audit trails. Okta separately flagged that this is part of a broader positioning as an AI access governance layer, with kill-switch capabilities for rogue agents and short-lived token scoping for agent-to-app connections.

The feature implements an open extension to the MCP authorization specification, meaning custom-built connectors can adopt the same behavior without waiting for Anthropic to certify them. That extensibility is the detail practitioners should note: organizations with proprietary internal tools can wire them into the same centralized governance model without going through a formal partner program.

First covered in Vol. I, No. 22. As of June 19, Day 7 of the suspension, no official restoration announcement has appeared from Anthropic or the Commerce Department. President Trump commented from the G7 summit in Évian-les-Bains on June 18 that negotiations with Anthropic are "going fine," marking the first direct presidential statement on the ban. A separate proposal to grant the UK an exemption from the directive has collapsed, narrowing the near-term path to any partial restoration.

Anthropic updated its privacy policy (effective July 8) to include government-issued ID and biometric data collection, a detail interpreted by observers as groundwork for a US-citizens-only restoration that would not require lifting the directive entirely. White House AI adviser David Sacks stated on June 14 that the administration's goal is for Anthropic to patch the jailbreak vulnerability so "Fable can go back online." Anthropic's technical rebuttal of the government's jailbreak assessment has not been published. Prediction market Polymarket priced restoration by July 1 at 75% as of June 14; current odds have not been independently verified.

In the meantime, the competitive consequences are becoming visible. Open-weight coding models from Cohere, Moonshot, and Zhipu moved from theoretical alternatives to actively evaluated fallbacks at enterprise accounts outside the US within the same week of the ban, per The New Stack's reporting. Fable 5 holds the top rank on the DeepSWE benchmark (70% PASS@1, per Datacurve), meaning the model remains offline at the moment of its strongest independent validation.

Published June 18 by OpenAI in collaboration with Molecule.one, the project used GPT-5.4 in conjunction with Molecule.one's Maria AI platform and its purpose-built micro-liter scale high-throughput experimentation (HTE) lab. The AI system chose the research area, generated and rated proposals, and directed the experiments. The full process from literature review to validated result took approximately 2.5 months, with an additional half-month for write-up. A preprint is available; it has not yet undergone formal peer review.

The specific reaction was Chan-Lam coupling involving primary sulfonamides, a variant that has historically produced yields too low for practical pharmaceutical use. GPT-5.4 proposed an unexpected modification; that proposal was then tested across 10,080 reactions in the automated lab and confirmed in manual replication. OpenAI describes the proposal as "unexpected" relative to what the literature had converged on. The result is notable because the model was not retrieving a known answer from training data but proposing something novel that then survived physical experimental testing.

The broader significance, per Molecule.one's co-announcement, is the closed loop: an AI system that can pick a research direction, generate hypotheses, run experiments, and surface a validated finding is a different category from an AI system that predicts outcomes from existing experimental data. The caveat is that this was a single research target, and generalization across diverse drug-discovery contexts has not been demonstrated.