Squared: Agents in Production, Chips, and a Government Pause Button

OpenAI is reportedly releasing GPT-5.6 to a select group of partners rather than the public, after the Trump administration raised security concerns — confirmed by Altman in an internal Q&A and reported by The Information, The Verge and TechCrunch.

This is the first time a frontier release has been visibly staggered at government request. If your roadmap assumes the newest model lands on day one for everyone, stop assuming that. Model availability is now a political variable, not just a commercial one — build supplier flexibility into your AI plans rather than betting a product launch on a single frontier model arriving on schedule.

As Bruce Schneier and Simon Willison both flagged, a landmark German ruling declared Google's AI Overviews to be Google's own words — making it legally liable for false answers the system generates.

This is the question every executive deploying customer-facing AI has been quietly avoiding. If a court treats AI output as your statement, your chatbot's hallucination is your liability. Revisit who owns the risk when your agent gives a customer wrong pricing, advice, or a false promise — and make sure your contracts, disclaimers, and human-review gates reflect that you are on the hook, not the vendor.

OpenAI, AWS and NVIDIA all pushed agent-to-production content this week (agentic overlays on legacy systems, secure agent runtimes), Notion is killing its email app because users let agents run the inbox, and the top trending GitHub repos are almost all agent harnesses and orchestration loops.

The conversation has shifted from 'can agents do tasks' to 'how do we control them at scale'. But Ford's cautionary tale lands in the same week — it had to rehire engineers to fix mistakes its automated systems made. The lesson isn't 'don't automate', it's 'automate with verification'. Patronus AI raising $50M purely to stress-test agents tells you where the maturity gap is. If you're deploying agents, budget for the testing and oversight layer now, not after the incident.

Anthropic claims Alibaba used 25,000 accounts to extract capabilities from Claude across 28.8 million exchanges — and wants it punished. Separately, Claude is quietly winning paying consumers from ChatGPT.

Two signals in one. First, your model provider's IP and access policies are becoming a battleground that could affect terms and availability — single-vendor dependence is a real exposure. Second, the paid-consumer shift toward Claude is a reminder that 'ChatGPT won' is lazy thinking. Re-evaluate which model your teams actually use by task, not by brand habit.

OpenAI and Broadcom unveiled Jalapeño, a custom inference chip, while IBM claimed the first sub-1 nanometer transistor technology and ASML's $400M lithography machines underpin it all.

This is incremental for your next quarter but structural for your next three years. Custom inference silicon is how providers will cut their own costs and differentiate on speed and price — which eventually flows to what you pay and how fast your applications respond. No action needed today, but if your AI strategy assumes today's cost and latency curves are fixed, they aren't.