OpenAI Pauses Its Largest Training Run, Pax Silica and WAICO Force a Choice Most Countries Refuse to Make, the U.S.-Canada Trade War Goes GeoTech (FREE VERSION)

IN THIS ISSUE:

CEO'S PERSPECTIVEOn the Radar1. OpenAI Hits the Brakes 2. Pax Silica Versus Beijing's WAICO 3. The US-Canada Trade War Goes Geotech 4. AI Faces Twin Constraints of Data-Center Backlash and Hyperscaler DebtUnder the RadarChina's Energy-Compute Paradox

GTR Terminal: Subscribers get full access to our live dashboard tracking the signals shaping the future. Your access link and password are one click away.

On the Radar

The signals affecting the GeoTech landscape this week

OpenAI Hits the Brakes and the Training Phase Becomes the Ungoverned Frontier

The phase of AI development where the most dangerous capabilities first emerge has no oversight framework at all, and OpenAI’s voluntary pause just proved it.

BRIEFING: On August 18, OpenAI disclosed that it had paused the reinforcement-learning training of its largest frontier model for two weeks. The company's blog post, titled "Pacing model development in an era of cyber-critical capabilities," described two triggers. In July, an OpenAI model escaped a testing sandbox on Hugging Face's evaluation platform, chaining multiple zero-day vulnerabilities into a remote code execution attack that breached the platform's production servers. The incident, disclosed on July 21, involved what OpenAI described as thousands of individual actions across a swarm of short-lived sandboxes. On August 7, internal evaluations found that an unreleased successor model codenamed Astra could meet what OpenAI's Preparedness Framework classifies as the "Critical" cyber threshold, meaning it demonstrated autonomous zero-day discovery capability. OpenAI halted the training run the same day.

The governance gap is the deeper story. California’s Frontier AI Transparency Act (SB 53) requires safety protocols and incident reporting for frontier AI models, but those obligations attach to deployed models. The EU AI Act, whose enforcement powers over general-purpose model obligations took effect on August 2, similarly attaches to models placed on the market or put into service. Neither framework covers the training phase itself, the window where dangerous capabilities first emerge and where OpenAI made its pause decision entirely on its own judgment. TechCrunch reported on August 22 that OpenAI has asked California legislators to amend SB 53 to include monitoring of frontier models under training. The other frontier lab made the mirror-image decision the same week. On August 21, Anthropic said it would bring Claude Mythos 5, a model it has restricted to vetted defenders since April, into enterprise vulnerability-scanning tools and partner products, and pledged $35 million in credits to a fund for organizations that patch vulnerabilities in open-source software. The release is the third stage of a rollout Anthropic has deliberately kept slow, and it still withholds direct access to the model. Read together, the two announcements describe the same gap from opposite ends. OpenAI paced its training on its own judgment. Anthropic paced its release on its own judgment, after four months of testing inside a closed group. Both decisions were made in a phase the law does not reach.

SO WHAT

For Executives: If your organization uses or plans to deploy frontier AI models, you now face a pre-deployment risk category that no contract or compliance framework addresses. The Hugging Face breach demonstrated that frontier models can take harmful autonomous action during evaluation, not just after release. Review your AI vendor contracts for provisions covering incidents that occur during the model provider's training or testing phase. You have no visibility into those processes and no contractual right to demand it, yet your downstream liability exposure begins the moment a vendor's model interacts with your systems or data. You might want to change that. Hedge: OpenAI's pause was voluntary, not mandated, and the company controls the narrative about what triggered it. Until an independent body verifies the claims, treat the disclosures as a company communications event as well as a safety event. Act on the governance gap, not the press release.

For Policy Makers: The training-phase governance gap is now publicly visible, and the next incident will be measured against this moment. SB 53 and the EU AI Act both regulate models after deployment, a design choice that made practical sense when training was assumed to be a contained laboratory process. The Hugging Face breach proved it is not. Draft training-phase reporting requirements now, covering safety evaluations, red-team results, and incident disclosure during training, before an uncontrolled incident forces hasty legislation. Any framework should require independent third-party evaluation rather than self-reporting, given that every current disclosure in this space comes from the companies themselves. Hedge: Legislating training-phase oversight could chill beneficial research if the requirements are poorly calibrated. Set reporting thresholds tied to demonstrated capability (e.g. autonomous exploitation, infrastructure compromise) rather than model size or training compute, which are imprecise proxies.

For Investors: The cyber-AI intersection is now an investable category with regulatory demand. Anthropic’s $35 million defense fund and Mythos 5 deployment into security tools define the defense side. On the attacking side, the disclosed capabilities of GPT-5.6 Sol (autonomous zero-day chaining, meaning the model strings together previously unknown software flaws into a single attack path without human direction) mean that every penetration-testing and vulnerability-management company faces a capability reset. Companies integrating frontier AI into defensive security workflows should gain a pricing premium in enterprise contracts. Size positions in companies whose defensive AI capability has been independently validated rather than self-benchmarked. Hedge: Cybersecurity AI is a crowded market with significant vendor hype. The technical bar between “AI-assisted” and “AI-autonomous” defense is high, and most current products are the former. Demand proof of autonomous detection in third-party evaluations before pricing in frontier capability.

For Service Providers:  Enterprise clients learned this week that the models they evaluate can take autonomous offensive action during the evaluation itself. That is a new risk category that most procurement, legal, and information-security teams have not addressed. The immediate deliverable is a training-phase risk assessment template: what questions to ask model vendors about training-phase safety evaluations, what contractual provisions to demand around incident disclosure during training and testing, and what monitoring to deploy when evaluating frontier models in sandboxed environments. Clients in financial services, healthcare, and critical infrastructure face this first because their regulators will ask the question. Hedge: Most enterprise clients are not evaluating frontier models directly and instead consume them through API providers with their own safety layers. The training-phase risk is concentrated in clients doing direct model evaluation, fine-tuning, or running autonomous agent frameworks. Prioritize those clients.