Week Ending September 11, 2026

Research window: September 5–11, 2026

1. Newsworthy

AI misuse is no longer mostly hypothetical; Anthropic is documenting real attempts to use capable models for weapons, espionage and biological research. Anthropic’s latest threat-intelligence report describes attempts involving cyber operations, surveillance, conventional weapons development, biological misuse and illicit model copying. Particularly important: Anthropic says some sophisticated dual-use biological work passed its normal safety classifier, strengthening the case for tighter access controls around the most powerful capabilities. Source

The 25th anniversary of 9/11 revived an old warning—“failure of imagination”—for an AI age. At a September 11 U.N. Security Council session, briefers highlighted terrorists’ growing use of AI, drones and encrypted platforms for recruitment and attack planning. A House Homeland Security anniversary review likewise identified AI and cybersecurity as part of today’s changed threat landscape. The connection is worth noticing: institutions built after 9/11 are now asking whether emerging technologies could create threats that look very different from those of 2001. Source

Congress is beginning to discuss an actual legal duty to build advanced AI safely. Bipartisan Senate negotiations would reportedly require frontier-model developers to exercise a “duty of care” against major known dangers, potentially including biological, nuclear and other catastrophic risks. The proposal could also give the federal government authority to temporarily block release of a model judged dangerously unsafe, subject to court challenge. It is far from enacted, but this is considerably more concrete than another voluntary safety pledge. Source

California is building an AI-auditing industry into law. Governor Gavin Newsom signed legislation creating standards for independent AI assessments and a registry for AI auditors. A day later, California enacted another package requiring stronger child-safety controls for companion chatbots, including independent audits and annual risk assessments. We have moved beyond simply telling companies to “use AI responsibly”; California is creating machinery for checking whether they do. Source

Schools are starting to put AI safeguards into contracts rather than policy statements. Microsoft, the American Federation of Teachers and the United Federation of Teachers agreed on protections that school districts can incorporate directly into Microsoft contracts. Among them: student and teacher data cannot be used to train models, schools retain control over data, high-risk decisions require human review, and families must receive plain-English explanations of AI use. Because these provisions are contractually enforceable, this is materially different from a set of suggested principles. Source

Wipro says AI has created productive capacity equal to about 20,000 employees—and it redeployed the people instead of simply eliminating them. Wipro has about 243,000 employees, and more than 100,000 have received advanced AI training. Its CTO says AI initiatives have freed capacity equivalent to the output of roughly 20,000 employees, who were redeployed to other work. The important business question may be less “How many jobs did AI replace?” and more “How much more can the same workforce now accomplish?” Source

A new national study suggests generative-AI adoption at work is broad but still surprisingly shallow. Researchers using a nationally representative survey found GenAI already appears across many occupations and tasks, but fewer than half of workers performing most individual tasks actually use it. That means occupation-level predictions can hide enormous differences between people doing essentially the same work. Who adopts AI—and how—may matter nearly as much as which jobs are theoretically exposed to it. Source

The New York Times copyright case may finally force a major U.S. court to answer whether training AI on copyrighted material is fair use. The Times, authors, OpenAI and Microsoft have filed competing requests for summary judgment in the consolidated litigation. The core question is no longer whether copyrighted works were used; it is whether using them to extract patterns for AI training is legally transformative or improperly creates competing products from unlicensed material. The ruling could reshape model economics. Source

Model distillation is turning into an international technology and privacy dispute. The U.S. accused several Chinese AI companies of industrial-scale copying of American models through distillation—using one model’s outputs to train another. Anthropic separately says it detected enormous unauthorized campaigns and alleges that some Chinese labs routed customer conversations through Claude, potentially exposing company data, credentials and personal information while harvesting Claude’s capabilities. China rejects the broader U.S. accusations. Source

AI data centers are beginning to look like strategic infrastructure that may need physical military-style protection. Reuters reports that the UAE is redesigning a planned 5-gigawatt AI campus after Iranian attacks damaged regional technology infrastructure, considering geographically dispersed facilities, hardened structures and underground components. Separately, the Pentagon is reportedly considering a $5 billion loan related to the U.S. AI data-center supply chain. Compute is becoming an economic asset, an energy issue—and increasingly a national-security asset. Source

A courtroom mistake provides a remarkably simple reminder of what responsible AI use still requires: check the answer. The New Mexico Supreme Court fined an attorney $5,000 after a brief included fabricated witness and police testimony generated by ChatGPT. This was not an exotic model-safety failure. It was an ordinary professional failing—using AI output without verification—which may ultimately be the more common risk facing most businesses. Source

2. Signals

AI governance is moving from principles to things you can actually audit, contract and insure.

This may be the clearest change of the week.

The Senate is discussing a legal duty of care for frontier developers. California is creating standards and registries for independent AI auditors. Its child-safety law requires risk assessments and independent audits. Microsoft and the teachers unions created AI protections that become enforceable contractual terms when school districts adopt them. And Anthropic’s misuse report argues that some powerful capabilities may need tighter, identity-based access.

Put those together and something important is happening.

For several years, “responsible AI” often meant: We have principles.

The emerging version looks more like: Show me your controls. Show me who had access. Show me the logs. Show me your testing. Show me the independent audit. Show me what happens when something goes wrong.

Those are much more familiar business questions. And they are exactly the questions a customer, regulator, auditor—or insurer—can price.

Signal worth keeping: the development of AI-audit standards may become just as economically important as the development of AI models.

The workforce story is becoming less about “Which jobs disappear?” and more about “Why do two people doing the same job get very different results?”

The new task-level research says something I think matters. AI use is widespread across occupations, but adoption within individual tasks is often below 50%. In other words, two accountants, programmers, salespeople or analysts doing similar work may currently use AI very differently.

Now put that beside Wipro. Wipro says AI has created capacity equal to about 20,000 workers. But it says those people were redeployed, more than 100,000 employees have received advanced AI training, and some engineers are increasingly managing agents rather than personally performing every step.

My inference is that the productivity gap within occupations could become more important before we see massive whole-occupation replacement.

Signal worth keeping: watch for evidence of widening productivity differences between workers with similar jobs, not merely employment changes between occupations.

“Human + AI” is beginning to mean management, not just assistance.

We’ve been talking about AI as a coworker. Wipro gives that idea a more concrete form. Its CTO described engineers who may manage “a bunch of agents” while working on other tasks. Meanwhile, the education agreement requires human oversight of consequential AI decisions, and California is requiring formal risk controls around systems interacting with children.

The important human skill isn’t necessarily: Can you get AI to produce something?

Increasingly it may be: Can you decide what to delegate? Can you define success? Can you recognize failure? Can you review several AI workers efficiently? Can you intervene at the right moment?

That sounds very much like management.

Signal worth keeping: “AI management” may become a basic professional skill rather than a specialized technical role.

Twenty-five years after 9/11, “failure of imagination” has a new technological dimension.

The original 9/11 Commission famously described a failure of imagination. On the 25th anniversary, that phrase resurfaced as national-security institutions looked at a threat environment that now includes cyberattacks, AI, biotechnology, drones and autonomous systems.

The U.N. Security Council heard that terrorist actors are increasingly exploiting AI, drones and encrypted platforms. The House Homeland Security Committee separately highlighted the double-edged nature of AI: it can help defenders find vulnerabilities and analyze enormous amounts of security data, but the same capabilities can help malicious actors.

Now connect that with Anthropic’s threat report from the same week. Anthropic says it has already disrupted attempts to use Claude for cyber operations, surveillance, biological misuse and conventional weapons work.

The point is not that AI creates another 9/11. That would be speculation. The useful connection is narrower: security institutions are explicitly asking whether they are adequately imagining threats made possible or amplified by technologies that barely existed in 2001.

Signal worth keeping: the next major security failure may come not from ignoring a known threat, but from applying yesterday’s threat model to a technology that changes what a small group—or even one person—can do.

Model distillation is becoming a question about ownership, privacy and national power all at once.

Distillation itself is not new or inherently improper. AI developers routinely use one model to help train another.

What changed is scale and context. The U.S. government is accusing Chinese firms of systematically extracting American AI capabilities. Anthropic says it identified large unauthorized campaigns and claims that some competitors routed actual customer interactions through Claude.

That creates three overlapping questions: intellectual property, privacy and geopolitics. We don’t have settled answers. But this looks like another area where our existing concepts were built for a world before generative AI.

Signal worth keeping: watch whether model-output rights and disclosure of downstream model providers begin appearing explicitly in contracts and privacy policies.

The AI infrastructure boom is turning into a national-security buildout.

Until recently, the constraints we kept watching were chips, electricity, financing and construction. Add physical security to the list.

The UAE is reportedly reconsidering the architecture of a five-gigawatt AI development because concentrated computing infrastructure can become a military target. The Pentagon is considering financing aimed at strengthening the domestic AI data-center supply chain.

My inference is that we should stop thinking of the largest AI facilities as simply oversized corporate data centers. They increasingly resemble power plants + semiconductor infrastructure + communications hubs + strategic computing reserves.

Countries may eventually treat them accordingly.

Signal worth keeping: geographic concentration of computing capacity may become an explicit national-security risk.

Verification may become one of the biggest hidden costs of AI.

The New Mexico attorney story is almost embarrassingly ordinary. ChatGPT generated false material. A professional did not check it. The professional—not the model—was sanctioned.

Now connect that with California creating AI auditors, Microsoft requiring human review for high-risk school decisions, Anthropic building threat monitoring, and employers trying to calculate actual business outcomes rather than raw AI output.

We often describe AI savings as: old human effort – new AI effort = savings.

But that leaves something out: verification.

As AI produces more work, somebody has to decide which work requires checking, how much checking is enough, and who is qualified to do it.

A better business equation may be: AI productivity gain = work automated – supervision and verification required – cost of errors.

Signal worth keeping: as AI output scales, verification may become a measurable operating cost rather than an informal human habit.

3. Summary

Okay, after everything I read this week, what should you take away from it?

I keep coming back to one number:

20,000 workers.

Wipro says its use of AI has created productivity gains equal to the output of about 20,000 employees. But here’s the part I think matters most: it didn’t eliminate those 20,000 people. It redeployed them. Wipro has about 243,000 employees overall, and more than 100,000 have received advanced AI training.

Think about what that means.

If Wipro’s estimate is right, AI didn’t simply replace human work. It created the equivalent of roughly 20,000 additional workers’ worth of productive capacity, while the existing employees could be moved to other work.

That’s a very different picture of AI and employment.

We’ve spent years asking whether AI will replace people. Wipro gives us a real-world example of another possibility:

What if AI makes the same workforce capable of doing substantially more?

That doesn’t mean layoffs won’t happen. They already have elsewhere in the industry. And Wipro still has to turn this additional capacity into revenue and better margins.

But it changes the question.

The important measure may not be how many employees AI replaces.

It may be how much additional productive capacity a company can create from the people it already has—and what it does with that capacity.

That may turn out to be one of the most important business stories in AI.

4. Reviewed for This Brief

Company Announcements, Reports and Primary Sources

• Anthropic — “Detecting and countering misuse of AI: September 2026” — September 10, 2026. Source
• Anthropic — “Measuring tactical intelligence targeting and conventional weapons capabilities of AI models” — September 10, 2026. Source
• Anthropic — “An alignment assessment of recent cybersecurity incidents” — September 9, 2026. Source
• Microsoft / AFT / UFT — “National AI Safety & Privacy Standard for schools” — September 9, 2026. Source
• California Governor’s Office — “First-in-the-nation AI safeguards” — September 9, 2026. Source
• California Governor’s Office — “Child-safety chatbot and social-media laws” — September 10, 2026. Source
• United Nations Security Council Counter-Terrorism Committee — “Security Council marks 25 years since 9/11 with counter-terrorism focus on AI and drones” — September 11, 2026. Source
• U.S. House Committee on Homeland Security — “25th Anniversary Report: National Security Improvements and the Domestic Threat Landscape 25 Years After 9/11/01” — September 10, 2026. Source

Research and Surveys

• Equitable Growth / Bick, Blandin, Deming and Schumacher — “What Work Does Generative AI Do?” — September 8, 2026. Source

Independent News and Analysis

• Reuters — “Wipro’s AI push frees capacity equivalent to 20,000 workers” — September 10, 2026. Source
• Reuters — “U.S. Senate negotiators consider requiring AI firms to mitigate known major risks” — September 11, 2026. Source
• Reuters — “U.S. accuses Chinese AI firms of malicious copying of AI technology” — September 8, 2026. Source
• Reuters — “OpenAI, New York Times case tees up key test of AI training under copyright law” — September 8, 2026. Source
• Reuters — “UAE revises AI data center plan after Iranian attacks” — September 11, 2026. Source
• Reuters — “Pentagon in talks to lend $5 billion to AI cloud startup Fluidstack” — September 10, 2026. Source
• Reuters — “Lawyer cites fake witnesses in murder case and blames ChatGPT” — September 11, 2026. Source

A Note About This Research Brief

Granddaddy’s Research is prepared by Two Trees Group, LLC for informational and educational purposes. We use human editorial judgment supported by AI-assisted research, analysis, and drafting to review a large body of material and identify developments we believe are worth your attention. AI can make mistakes, and so can we.

Whenever practical, we provide links or references to original sources so you can review the underlying material yourself and verify information that is important to you. Our summaries, commentary, analysis, Signals, assumptions, and conclusions represent our interpretation of the material reviewed. They are not a substitute for professional advice.

© 2026 Two Trees Group, LLC. Original summaries, commentary, analysis, Signals, and other original editorial content are protected by applicable copyright law. Third-party source materials remain the property of their respective owners.

Grace and Peace,

Scott Walker

GranddaddyCant.com