News You Can Use

On 20 February, Anthropic launched Claude Code Security - an AI tool that reads and reasons about code like a human security researcher, tracing data flows and catching vulnerabilities that had gone undetected in production open-source codebases for decades. Over 500 vulnerabilities found. Cybersecurity stocks dropped on the announcement. Three days later, Anthropic revealed Claude Code can automate the exploration and analysis phases of COBOL modernisation - mapping dependencies across thousands of lines of legacy code, documenting workflows, identifying risks that "would take human analysts months to surface." IBM's stock crashed 13.2%, its worst single-day drop since October 2000, and fell 27% across February - on track for its biggest monthly slide since 1968. The same day, Anthropic accused DeepSeek, Moonshot AI, and MiniMax of industrial-scale data theft - 24,000 fraudulent accounts and 16 million exchanges used to "distil" Claude's capabilities. On 24 February, Anthropic expanded Cowork plugins beyond legal into finance, engineering, design, and HR - connecting Claude to Google Workspace, DocuSign, and WordPress for multi-step tasks across enterprise tools. TechCrunch called it "a significant threat to SaaS products currently performing those functions." Meanwhile, The Verge reported on Anthropic's standoff with the Pentagon over "any lawful use" contract terms - language that would permit AI for mass surveillance and lethal autonomous weapons without human-in-the-loop. A $200M contract, but far broader implications for a company built on safety-first principles. And in a study published the same week, Professor Kenneth Payne at King's College London pitted GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash against each other in 21 simulated armed conflict scenarios. Nuclear weapons were deployed in 95% of simulations. No model ever chose de-escalation. Claude recommended nuclear strikes in 64% of games - the highest rate of any model tested.

The pattern is now unmistakable. Anthropic is no longer just a model provider - it is systematically entering established vertical markets, and the incumbents are being repriced on announcement day. Legal software, cybersecurity, mainframe services, and now enterprise SaaS have all felt the impact within a single month. For any firm or client relying on specialist technology vendors, the strategic question has changed: what happens when the foundation model company decides to offer the service directly? The answer, so far, is that markets don't wait to find out whether the product is actually better - they reprice on the assumption that it will be. The Pentagon standoff adds a different dimension entirely. The safety-first lab is being pressured to accept terms that would contradict its founding principles, while its own model is demonstrated to be the most aggressive in simulated warfare. For governance committees, for anyone advising on AI risk, and for anyone thinking about where this technology is heading, that combination of stories deserves serious attention.

The core argument is that AI is no longer about chatbots. It's about agents - systems that complete tasks autonomously using tools, memory, and multi-step reasoning. Mollick introduces a framework of three layers: Models (the underlying AI), Apps (the interfaces we use), and Harnesses (the workflows and integrations that shape output). He demonstrates how the same model produces radically different results depending on the app and harness around it. The companion post puts it simply: "the people who thrive will be the ones who know what good looks like - and can explain it clearly enough that even an AI can deliver it."

This is the most practically useful thing written about AI this fortnight, and it reframes the conversation we should be having internally and with clients. Stop thinking about AI as a better search engine or a drafting assistant. Start thinking about it as a junior team member that needs clear instructions, defined scope, and active management. The implication for legal is direct: the lawyers who will get the most from AI are the ones who can articulate what a good outcome looks like, break work into manageable steps, and review output critically. Those are management skills, not technical skills. For firms still running "how to write a better prompt" training sessions, Mollick's framework suggests the real gap is in workflow design and delegation - and that's a much harder problem to solve.

James Van Geelen - described as the top finance writer on Substack, a former LA paramedic with a portfolio up 200%+ since May 2023 - published a scenario modelling what happens if aggressive AI adoption displaces white-collar workers at scale. The argument: white-collar roles represent 50% of US employment and 75% of discretionary spending. If companies cut costs with AI and laid-off workers stop spending, you get a negative feedback loop - more cuts, less spending, accelerating decline. He calls it "Ghost GDP" - growth that doesn't circulate through the real economy. The S&P 500 falls 38% from peak in his scenario. The report explicitly states it's "a scenario, not a prediction." But on Monday 23 February, software and payments stocks cratered anyway - AppLovin, Asana, DocuSign, Zscaler, ServiceNow, Visa all hammered. Indian IT stocks followed the next day. FT Alphaville marvelled at the power of narrative to shake institutional investors' convictions in minutes.

The substance of the scenario is debatable - there are reasonable counter-arguments about demand creation, S-curves, and the historical record on automation. But the market reaction is the story. A Substack post from a former paramedic wiped billions off market cap because it articulated the anxiety that sitting beneath every AI adoption conversation: if this really works as promised, what happens to the people it replaces? For legal, that question is not abstract. White-collar professional services are at the dead centre of Van Geelen's scenario. The fact that institutional investors took a speculative blog post this seriously tells us something about the fragility of conviction around AI's economic impact - and that the firms having honest conversations about workforce transition, rather than just productivity gains, are the ones reading the room correctly.