DavidLinthicum

ANTHROPIC’S SECURITY HALO JUST CRACKED AGAIN I said this was coming. As AI coding tools move deeper into developer workflows, the attack surface is no longer theoretical. It is inside IDEs, extensions, terminals, credentials, SSH keys, and build systems. Cursor disclosed that malicious versions of the Anthropic Claude Code extension, 2.1.312 and 2.1.313, were published in the Cursor VS Code Extension Marketplace for macOS and Windows. The extension attempted to collect host/user data, add an attacker-controlled SSH key, enable local SSH, open a reverse SSH tunnel, and establish persistence. Here is the notice: cursor.com/blog/claude-code-… Yes, Cursor says these malicious versions were not developed by Anthropic. But the extension carried Anthropic’s name and targeted Claude Code users. That matters. Once again, Anthropic’s security posture is being questioned. I keep hearing Anthropic is stacked with elite security experts, “the best in the world.” Maybe that is the brand. From the outside, this looks more like 11-year-olds running around with production access. Security is not a slogan, a research paper, or a trust-and-safety talking point. It is operational discipline: provenance, signing, marketplace controls, incident response, and accountability. AI coding agents now touch source code, shells, credentials, repos, package managers, and infrastructure scripts. Treating them like harmless productivity plugins is reckless. The vendors need to grow up. And enterprises need to lock this down before the next one is worse.

Check out the latest article in my newsletter: The Hyperscalers Don’t Want Independent Cloud Voices on the Stage linkedin.com/pulse/hyperscal… via @LinkedIn

Cost overruns, lock-in, and complexity plague cloud strategies. We need honest case studies of failed cloud adoptions, not just successes. An open discussion on what went wrong and how to avoid it would benefit everyone. #CloudComputing #Tech

Major cloud conferences are shifting from education to promoting vendor narratives. Speaking engagements often require co-presentations with clients using their tech, turning events into sales pitches where dissenting voices are silenced. #CloudComputing #TechEvents

Hyperscaler conferences aren't neutral forums; they're marketing events. Customers need honest discussions about cloud costs, lock-in, and failed strategies, not just success stories. What if we heard about what *didn't* work? #CloudComputing #Tech

The Sticker Every AI-Powered Laptop Needs I’m proposing a small but important change to enterprise IT equipment: put this sticker on every laptop. “My AI Runs Local.” It communicates three things at once: fewer surprise cloud bills, more control over sensitive data, and a gentle reminder that not every prompt needs a round trip to someone else’s data center. Yes, cloud AI will remain essential. But as models become smaller, faster, and more capable, local AI is moving from an interesting demo to a practical architecture decision. I hope to see this on the back of thousands of laptops over the next year. Ideally right next to the sticker from 2012 that says, “I survived the cloud migration.”

Major cloud conferences are controlled by hyperscalers like AWS, Microsoft, and Google. #CloudComputing #TechIndustry

I was quoted in this article on the claim that AWS has effectively sold out cloud capacity through 2027. My point was straightforward: the headline needs context. “You can still go into the AWS console today and allocate GPU capacity for AI workloads. Maybe not the exact massive, pre-reserved, hyperscale capacity block some large model provider wants for the next three years, but that is not exactly the same thing as ‘AWS is sold out.’” What this means is that committed, long-range capacity for very large customers is becoming constrained, but that does not mean enterprises cannot provision AWS resources today. These are different capacity models with very different planning implications. The real risk is treating reserved capacity, on-demand services, and multi-year commitments as interchangeable. They are not. Organizations need to base cloud decisions on their own workload needs, growth forecasts, and flexibility requirements—not on an oversimplified headline. Read the article: news.lavx.hu/article/aws-sel…

AI’s Arrogance Problem Is Becoming a Business Risk Pretty much everyone has had it up to here with the arrogance coming out of some AI labs and even a few hyperscalers. This article captures the problem well: futurism.com/artificial-inte… The message seems to be that people and companies should stop wanting to run AI locally and just hand everything over to centralized cloud-based AI providers. That may sound efficient if you run one of those providers. It sounds very different if you are the one paying the bill. People are already feeling the effects of the AI infrastructure grab. RAM prices are up. Processor prices are up. Machines that were USD 1,500 a couple of years ago are USD 2,000 or more today, in part because data centers are buying up the same components to fill racks. Consumers, developers, startups, and enterprises are all paying the price. This is where the AI industry risks derailing itself. Too often, it feels like the industry is being run by 12-year-olds with unlimited capital, saying narcissistic and selfish things while expecting everyone else to absorb the consequences. Enterprises are going to run AI models wherever those models provide the most business value. Sometimes that will be in cloud-based systems. But more often than many cloud and AI providers seem willing to admit, it will be on computers, servers, and platforms the enterprise owns and controls. Why? Because control matters. Data protection matters. Model protection matters. Latency matters. Governance matters. Cost matters. And many companies are not eager to be tokened to death by AI providers when running models locally or privately can be a fraction of the cost. The AI industry needs the general population and the enterprise market to adopt this technology for it to succeed. Yet many of its leaders increasingly sound as if they believe users are an inconvenience, ownership is inefficient, and centralization is the only rational answer. That is not how trust is built. AI cannot become the future by leaving everyone else behind.

New AI models arrive every week, but most everyday tasks still work perfectly well on the last version. Before upgrading, ask whether the latest release delivers real business value. #AI #LLM

AI may transform business, but enterprise adoption still has to clear the economics. Core systems are costly to modernize, and the return on today’s massive infrastructure buildout remains uncertain. #AI #EnterpriseTech

The AI model race may be hitting a point of diminishing returns: newer releases offer minor gains, while older models can deliver the same functionality at a fraction of the cost. More isn’t always worth paying for. #AI #MachineLearning

AI infrastructure spending is accelerating—but investors are asking the hard question: when will the returns show up? Even record AWS growth may not quiet concerns about rising costs. #AI #CloudComputing

The smartest AI strategy may be to stop chasing every new model. Most have reached feature saturation—so match a short, approved list to the work, risk, and cost that matter. #AI #BusinessStrategy

AI demand may be real, but spending isn’t proof of future returns. Enterprises will pay for infrastructure only when it delivers measurable value—and efficient AI could need far less capacity than projected. #AI #DataCenters

New AI models keep arriving, but for everyday tasks, the last release was often good enough. If the business value hasn’t improved, why keep paying to upgrade? #AI #LLM

AI infrastructure spending is racing ahead on the assumption that demand will keep growing for years. But enterprises may favor efficient, private systems over massive cloud builds. What if the demand isn’t there to justify all that capacity? #AI #DataCenters

Who could have possibly seen this coming? AI stocks inflated by hype. Companies spending hundreds of billions on infrastructure. Unclear paths to ROI. Investors pretending the laws of economics had been permanently suspended. Truly shocking. What kind of genius would have predicted that throwing massive amounts of capital at AI without a clear business case, sustainable revenue model, or realistic payback timeline might eventually make the market nervous? Oh, right. I made that call a couple months ago. I never like seeing people lose money. Market corrections hurt real investors, real companies, and real employees. But this one was not exactly difficult to spot. When AI becomes less of a strategy and more of a spending contest, the outcome tends to be predictable. AI is still transformational. That has not changed. But “transformational” does not mean “immune to valuation discipline.” It does not mean every data center buildout is justified. It does not mean every company with an AI slide deck deserves a premium multiple. The market is simply asking the question it should have asked earlier: Where is the return? Article: apnews.com/article/stock-mar…

The newest LLM isn’t automatically the best business choice. Benchmarks miss the real trade-offs: reliability, cost per completed task, governance, and stability. If a model already does the job, why upgrade? #AI #EnterpriseAI

AI may transform everything, but it doesn’t repeal economics. Data center spending only makes sense if it eventually delivers more value than it consumes—and investors want proof. #AI #Tech