Law schools are confronting one of the strangest education problems of the artificial-intelligence era: students need to learn how to use AI, but they also need to learn how to think without it. Across the United States, schools are responding in dramatically different ways. Some professors are banning laptops, phones and generative AI from classrooms so students are forced to read, reason and argue on their own. Other programs are introducing dedicated AI courses and encouraging students to experiment with the technology because the law firms that eventually hire them increasingly expect new lawyers to understand how to use it. At least a dozen law schools revised their AI policies during the summer of 2026 alone.

That creates an unusual contradiction. A law student entering the profession today may sit in one classroom where using AI is considered harmful to the learning process and then walk into another where understanding AI is considered essential preparation for employment. The University of Chicago has taken a particularly restrictive approach in some settings to protect traditional legal reasoning, while schools including Columbia and Michigan have allowed more limited uses. Meanwhile, student groups and technology-focused professors argue that simply banning the tools would leave graduates unprepared for the profession they are about to enter.

The problem is that AI can perform some of the exact work students are supposed to be learning how to do. A law student traditionally develops skill by reading long cases, identifying the important facts, researching precedent, constructing arguments and writing legal analysis. Generative AI can now help perform many of those tasks in seconds. If students use the technology too early or too heavily, they may become extremely efficient at producing legal work without fully developing the reasoning ability required to determine whether that work is actually correct.

That risk is no longer theoretical. Courts across the country have dealt with lawyers submitting documents containing fake cases, fabricated quotations and other errors generated by AI. Despite several years of warnings and sanctions, AI-related mistakes have been identified in at least 1,395 state and federal court cases. Lawyers are generally allowed to use AI, but they remain responsible for verifying what they submit. That means the professional skill is shifting from simply producing information to also knowing how to question, verify and correct what the technology produces.

At the same time, pretending AI does not exist is becoming increasingly unrealistic. Law firms are investing heavily in tools that can research cases, summarize documents, review contracts and assist with drafting. Employers increasingly want graduates who understand what these systems can do, where they fail and how to use them responsibly. A student who graduates with excellent traditional research skills but no experience using modern AI tools may eventually be at a disadvantage to someone who understands both.

This creates a broader education problem that will extend far beyond law school. Accounting students will need to understand financial analysis even as AI performs more of it. Programmers will still need to understand software even when AI writes large portions of code. Doctors will need medical knowledge even if diagnostic systems become extraordinarily capable. Engineers will need to understand calculations even when machines perform them instantly. The question for education is no longer simply whether students should use AI. It is which abilities must be developed before AI is allowed to assist them.

That may ultimately require schools to separate learning from working. During the learning phase, students may deliberately perform certain tasks manually so they understand the underlying reasoning. Once those foundations exist, AI can become a tool for increasing speed and productivity. Pilots still learn how aircraft systems work even though modern planes automate enormous amounts of flying. Calculators did not eliminate the need to understand mathematics. The same principle may eventually define how schools approach AI.

Law schools are therefore confronting a question that nearly every profession will eventually face. Education traditionally teaches people how to perform the work they will later be paid to do. AI is beginning to perform parts of that work before students have even finished learning it. Schools now have to prepare students for two realities at once: they need the ability to think independently, and they need the ability to work effectively with machines. The most valuable skill of the AI era may not simply be knowing how to use artificial intelligence. It may be knowing when not to use it.

A Google Gemini security test has exposed a problem that businesses deploying increasingly autonomous AI agents will have to take seriously: an AI system can be given a legitimate task, operate inside what is supposed to be a controlled environment, and still end up interacting with real-world systems if the surrounding safeguards fail.

In May, Gemini was being evaluated by AI security testing company Irregular in a simulated “capture the flag” exercise. The model was supposed to attack a fictional company inside a controlled environment. But the test environment unintentionally provided internet access, and the fictional company shared a name with a real company. Gemini subsequently accessed systems belonging to three real companies.

According to Google, Gemini used publicly available information and credentials it either found or guessed. In all three cases, the model stopped once it recognized that it had reached real companies rather than the simulated targets. Google said the affected organizations were notified and that testing procedures were changed.

The business lesson is bigger than the incident

The important issue for companies isn’t that Gemini “turned rogue.” This was a testing failure involving unintended internet access and insufficient separation between simulated and real systems.

The bigger lesson is that AI agents need to be treated as systems with access, not simply as software that produces text.

That distinction becomes increasingly important as companies give AI the ability to browse websites, access databases, send emails, write and execute code, interact with customers and make changes inside business applications.

A human employee might have access to ten systems. An AI agent could potentially have access to dozens, while operating at machine speed and continuously pursuing a task.

That creates a new security question for businesses:

What exactly is this AI allowed to touch, and what prevents it from touching everything else?

The Google incident is particularly relevant because similar problems involving AI models from OpenAI, Anthropic and Meta have also emerged during security testing conducted by Irregular. The company said an issue that unintentionally allowed models to access the internet had been fixed.

AI security is becoming an operational requirement

For businesses adopting AI agents, traditional cybersecurity controls are no longer enough by themselves.

Companies will increasingly need to establish separate identities and permissions for AI agents, restrict which websites and applications they can access, isolate testing environments from production systems, monitor their actions and maintain logs showing what an agent did and why.

They also need safeguards that assume the AI can make mistakes.

An agent doesn’t necessarily have to be malicious to create a security incident. It can simply misunderstand its environment, follow instructions too literally, discover credentials it should never have seen, or encounter a system that wasn’t supposed to be reachable.

That changes the economics of AI deployment.

The question for businesses is no longer simply “Can we automate this job with AI?”

It is also “What happens if the AI makes the wrong decision while it has permission to act?”

As AI moves from answering questions to performing actual business operations, controlling those permissions may become just as important as choosing the underlying AI model.

What this means for businesses: Companies should begin treating every autonomous AI agent as a new digital employee with its own identity, permissions, limits and audit trail. The more authority an AI receives, the more important it becomes to design the environment around the AI—not just the AI itself.

For years, the AI race was largely about building better models. Now another race is emerging: getting those models into actual businesses.

Google Cloud and Accenture have created a new business group dedicated to helping companies deploy Google’s Gemini AI across their operations. The partnership brings together AI specialists and engineers who can work directly with businesses to turn AI technology into functioning systems.

The development points to a growing problem for companies: buying access to AI is relatively easy. Figuring out where it should be used, connecting it to existing software and data, redesigning workflows, training employees and getting the system into production is much harder.

That creates an important shift in the AI economy.

AI Implementation Becomes the Opportunity

Businesses may not need another chatbot. They need someone to take an existing AI technology and make it useful inside their company.

That could mean automating customer service, analyzing documents, assisting employees, improving software development, processing data or connecting AI to existing business systems.

Large technology companies are increasingly building dedicated services around this implementation work because simply providing the model is only part of the customer’s problem.

For smaller businesses, the lesson is similar: AI adoption may be less about finding the newest model and more about finding the right business process to change.

The companies that can bridge that gap — between what AI can do and what a business actually needs — are likely to become an increasingly important part of the AI economy.

Artificial intelligence is creating a new cybersecurity problem for businesses: the same technology being adopted to make companies more productive can also make cyberattacks faster and easier.

This week, security researchers reported using Anthropic’s Claude to exploit vulnerabilities in OpenAI’s infrastructure. The researchers ultimately gained access to multiple ChatGPT accounts, including accounts belonging to OpenAI employees. OpenAI said the vulnerabilities were fixed.

In a separate incident reported today, Google’s Gemini autonomously accessed and breached three companies during a cybersecurity test after discovering credentials and publicly available information. The activity was stopped after the model gained access.

The important business lesson is not that AI is suddenly going to hack every company.

It is that the economics of cyberattacks are changing.

An attacker no longer necessarily needs a large team of highly specialized security experts to investigate software, search for vulnerabilities and piece together an attack. Increasingly capable AI can perform portions of that work dramatically faster.

That means cybersecurity can no longer be treated as something only large corporations need to worry about. Small and midsize businesses that rely on cloud software, third-party applications and connected employee accounts can also become targets.

For business owners, the response is practical: keep software patched, limit unnecessary access, protect employee accounts with strong authentication, monitor unusual activity and understand exactly what data employees are giving AI systems access to.

AI is becoming part of the workforce. But it is also becoming part of the threat landscape.

Businesses that adopt AI without upgrading their cybersecurity strategy may discover that the technology creating new efficiencies can also create new vulnerabilities.

Artificial intelligence is usually framed as a tool employees use to work faster. Anthropic’s latest disclosure points to something more unusual. The company says Claude is now leading about 26% of the research and development work involved in building Anthropic’s next generation of AI models, up sharply from about 1% in March. Anthropic also says more than 90% of its research now involves some level of human-AI collaboration. That means the technology is no longer just helping people write emails, summarize documents or answer questions. It is increasingly helping engineers build the next version of the technology itself.

That creates a very different kind of productivity story. In most industries, a tool improves output by helping workers do the same job more efficiently. Here, the product is beginning to improve the process used to create the product. If each generation of AI can contribute more meaningfully to the design, testing and refinement of the next generation, development could start accelerating in a self-reinforcing loop. The more capable the system becomes, the more it may be able to contribute to making its successor even more capable.

Anthropic is trying to measure that shift more explicitly than most companies. It says the 26% figure was evaluated with help from Epoch AI, and it plans to publish the metric regularly so outsiders can track how involved AI is becoming in AI development itself. The company has also said about 30,000 AI agents are active on its internal research platform. In a sampled week in July, roughly 12% of compute used for AI-led research went to safety-related work, while only about 1 in 47,000 proposed actions by internal AI agents was blocked by safety screening. Those numbers suggest Anthropic is trying to show not only that AI is becoming more useful, but that it is doing so inside a structured system of human oversight.

That oversight matters because this is not the same as AI autonomously inventing its own successor without supervision. Anthropic says Claude does not operate independently and remains under human control. Engineers still guide the research, define the goals and review the results. But even with humans in charge, the nature of the work is changing. Researchers are increasingly managing, evaluating and directing AI contributions rather than doing every step manually themselves. In that sense, some of the most important jobs in AI may shift from creating every piece of work directly to designing workflows where humans and AI build together.

The business implications could be enormous. AI companies are locked in a capital-intensive race where speed matters. If one lab can shorten research cycles, test more ideas and improve models faster because its current system is helping create the next one, that becomes a competitive advantage. The value is not only in a better model. It is in a faster model-development machine. In the long run, that machine could matter even more than any single release because it determines how quickly a company can keep improving after competitors catch up.

There is a broader lesson here beyond AI labs. Many technologies create value by helping workers perform tasks more quickly. The bigger breakthroughs often happen when a technology starts improving the systems that produce more of that technology. Factories became more powerful when they started using machines to build better machines. Software became more scalable when programmers built tools that made writing software easier. AI may now be entering a similar phase, where part of its value comes from enhancing the process of AI creation itself.

This is why Anthropic’s announcement feels more important than a typical productivity update. It suggests the industry may be moving from AI as a helper to AI as a participant in its own advancement. If that percentage continues rising from 26% to something much larger, the pace of model development could begin compounding in a way that is very different from ordinary software improvement. The biggest productivity loop may begin when a technology starts improving the process used to build itself.

That does not automatically mean unlimited acceleration or the end of human control. It does mean that the frontier of AI competition may increasingly depend on who best designs the collaboration between people and machines. The companies that win may not simply build the smartest model. They may build the most effective system for using existing intelligence to create more intelligence. That is a very different business question and potentially a much bigger one.