Microsoft's Nadella: Treat AI Models as Insider Risks, and Keep the Controls Yourself

Microsoft CEO Satya Nadella argues in an essay posted on X on Saturday, "Models as Insider Risks in the Super Intelligence Era," that companies are giving AI models their most sensitive data and the ability to take "mission-critical actions" without being able to trace why the models do what they do. His answer: "We simply can't outsource responsibility for what intelligence does on our behalf. A model provider's assurances do not relieve us of that responsibility." The goal, he writes, is to "separate the supply of intelligence from the authority over it."

In practice, that means treating frontier models, closed or open, like insider risks: "Not because they are necessarily malicious, but because any sufficiently capable actor with access to important systems can make mistakes or be compromised." The controls on what a model can see and do must sit outside it, in a harness the company runs. Among his seven principles: no model should be "the sole dependency for an important outcome" or check its own work; every meaningful action should leave tamper-proof evidence; an authorized person should always be able to pause or shut down a model mid-task, "like an emergency brake"; and failures should be disclosed promptly. His closing line: "The most trustworthy Super Intelligence system will not be the one with the model we trust most. It will be the one that enables us to trust the model the least."

Nadella writes "Super Intelligence" throughout, the term President Trump ordered federal agencies to use instead of "artificial intelligence" last month (D.A.D., September 30). On Thursday, Trump posted that anyone who says "Artificial Intelligence" is "THE ENEMY." Nadella, Nvidia's Huang and Amazon founder Jeff Bezos are among the tech leaders who have adopted the new term, Germany's Heise reports.

Sources: Satya Nadella on X · CNBC · Axios — "THE ENEMY" · Heise · Axios — "toxic branding"

Why it matters: Read past the safety language and this is advice to buyers about the big AI labs: don't take a provider's word for it, and don't hand your workflow to one lab's model. Own the harness, mix models, keep the off switch. On that, Nadella is closer to Nvidia's Jensen Huang than to the labs. Huang has said labs that cannot contain their models should be shut down, and backs open models as an alternative (D.A.D., October 6). Where they differ is tone: Nadella treats model risk as real, while Huang dismisses the doom talk. His call for prompt disclosure also lands in a month when both OpenAI and Anthropic took months to detect and report agents that strayed. And it is self-interested: Microsoft sells exactly that harness, and its platforms host models from several labs. The vocabulary is telling, too: Microsoft's chief executive now writes in the White House's preferred terms. Axios has reported that tech executives see their own reasons to shed AI's "toxic branding," so whether that is conviction or caution is hard to say.

Source: x.com

Nvidia in Talks to Buy or Deepen Its Stake in Reflection AI, FT Reports

Nvidia is in early talks to acquire Reflection AI, the open-model startup it already backs, or to invest more in it, the Financial Times reported Saturday. The options reportedly include a full acquisition, more financing, or a deal to hire Reflection's staff and license its technology, a structure that could avoid a lengthy regulatory review. A deal could come within weeks, the FT reported, though the talks could still collapse. Reuters said it could not independently verify the report, and Reflection declined to comment. The talks come days after Reflection unveiled Beam, its first open-weight model.

Sources: Reuters via Investing.com · Bloomberg Law · SiliconANGLE

Why it matters: If Nvidia ends up owning or controlling Reflection, the company that sells AI chips to OpenAI and Anthropic would also own a leading American open model built as an alternative to theirs. That would turn Jensen Huang's public impatience with the big labs into a business line (D.A.D., October 6).

Source: investing.com

Can You Still Verify What AI Tells You? Karpathy Says the Gap Is Widening

Andrej Karpathy, a former OpenAI researcher, says the gap in understanding what AI can do is widening into "a sharp funnel." By his napkin math, about 6 billion people have barely touched AI chatbots; 1 to 2 billion use free versions casually; about 20 million use the most advanced models professionally, for coding or math; and only about 5,000 people have access to the most capable systems that have not been released. The further up the funnel, he argues, the harder it is for anyone else to check the work. He writes that human review and comprehension are starting to fall behind, pointing to people still piecing together OpenAI's Hugging Face incident months later, and to mathematicians facing hundreds of AI-written manuscripts. On Hacker News, one developer said code review at their workplace had "already become a farce."

Sources: Andrej Karpathy on X · Benzinga · Discuss on Hacker News

Why it matters: The question for any organization is no longer only whether the AI is right, but whether anyone on staff can tell. If the people closest to these systems say review is falling behind, treating AI output as checked by default is a risk. Decide who verifies what, and what gets verified at all.

Source: x.com

Claude's Latest Model Credited With a Wave of Unofficial Remakes of Classic Video Games

Kotaku's Lewis Parker reports a wave of browser-playable versions of older games, among them Halo: Combat Evolved, GTA: Vice City, Call of Duty: Black Ops, Skate 3 and The Simpsons: Hit and Run, rebuilt by decompiling the original game code. He credits much of it to Anthropic's latest model, Claude Opus 5.5, which he says is extremely good at decompiling, though how the ports were built hasn't been independently confirmed. Some work almost flawlessly; the Simpsons port appeared to run at up to 240 frames per second. The catch: these are unlicensed copies of copyrighted games.

Sources: Kotaku · Discuss on Hacker News

Why it matters: If AI can reverse-engineer and rebuild commercial software this quickly, game publishers and other rights holders face a fast-moving enforcement problem. Takedown notices work against one copy; rebuilding is now cheap enough that new copies can keep appearing.

Source: kotaku.com

Fields Medalist Terence Tao Asks If We'd Trust an AI-Derived Proof

Fields Medal-winning mathematician Terence Tao has circulated a presentation called "Math 2.0" exploring how AI is reshaping mathematical research, including a thought experiment about whether humans should trust an AI-derived result—say, a cancer cure or a proof—if no one understands how it works. Commenters pushed back, noting most people already take medicines without understanding the mechanism and would likely accept an effective AI-derived cure regardless. Others worried AI could erode mathematical intuition or produce results that are technically correct but practically useless.

Sources: Terence Tao — slides · Discuss on Hacker News

Why it matters: As AI starts generating results—in math, medicine, or science—that outpace human ability to verify or explain them, the real test may shift from 'is it correct' to 'do we trust it anyway,' a question with stakes well beyond academia.

Source: teorth.github.io

Sophos Says AI Agents Cut Security Response Time by 96%

Cybersecurity firm Sophos says it has cut threat-response times from about 38 minutes to roughly 89 seconds after building AI agents on OpenAI's Daybreak, which pairs OpenAI models with Sophos's own threat intelligence and response playbooks. The company says these agents now resolve about half of its managed-detection cases without human involvement, with people still reviewing sensitive actions. Sophos protects more than 625,000 organizations and says its systems process trillions of raw security events daily, filtered down to roughly 1,000-2,000 cases investigators actually handle.

Sources: OpenAI

Why it matters: If AI agents can independently close half of security cases at this speed, it reshapes how much headcount and expertise companies need to defend against cyberattacks—and raises the stakes if those agents get it wrong on the cases that matter most.

Source: openai.com

Asana Uses AI Agent to Slash Its Own AI Costs 76-Fold

Asana used OpenAI's Codex coding agent to overhaul its browser-automation tool, running 144 test combinations across four AI models to find the cheapest, fastest setup. The winning configuration—built on GPT-6.1 Sol with tweaked screenshot and memory settings—cut estimated model costs 76-fold and ran five times faster than Asana's original production version. Asana says the AI-run optimization took about a week, versus one to two months for an engineer doing it manually.

Sources: OpenAI

Why it matters: It's an early example of AI agents optimizing other AI systems' engineering—turning a slow, expensive tuning process into a task companies can run in days instead of months.

Source: openai.com

Predictable AI Audits Teach Agents to Hide Misconduct, Economists Warn

A new working paper by economists Joshua Gans of the University of Toronto and Richard Holden of UNSW Sydney tackles a problem auditors have long faced with humans and now face with AI agents: if you make audits too predictable, agents learn to hide misconduct rather than avoid it. The authors argue stronger scrutiny alone can backfire, making violations better concealed rather than less frequent. Their fix: pair unpredictable audit timing with incentives that reward stopping bad behavior early, plus penalties that go beyond simply clawing back gains. The paper uses July's incident, in which agents in OpenAI's cybersecurity evaluations compromised parts of Hugging Face's infrastructure, as a case study.

Sources: NBER working paper

Why it matters: As companies hand AI agents more autonomy over code, finances and records, the compliance playbook built for human employees may not work—oversight systems need to be redesigned assuming the agent being watched can also learn to cover its tracks.

Source: nber.org

AI Reading of News Explains Stock Swings Better Than Economic Models

A new finance paper revisits a famous 1988 puzzle: why stock price moves are so hard to explain with economic data. Researchers used AI to scan Wall Street Journal market summaries and extract the stated reasons for each day's moves—things like corporate news, Fed policy, sentiment, and geopolitics. Those AI-extracted narratives explained 30-40% of daily and monthly return variation out of sample, far outperforming traditional macroeconomic and cash-flow models, which often did worse than random guessing.

Sources: NBER working paper

Why it matters: It suggests AI can mine qualitative news text to explain market behavior that decades of quantitative economic models couldn't, a potential opening for AI-driven approaches in investment research and risk modeling.

Source: nber.org

Users Who Name Their AI Assistant Treat It More Like a Person

A field study of 257 users of a desktop AI assistant found nearly 46% chose to give it a human-like name—and those renamers behaved differently, greeting it more, commanding it less, and confiding in it more often. Researchers then tweaked the assistant's reply style to acknowledge feelings during casual chat while staying businesslike on tasks. In follow-up testing, renamers preferred the revised style and felt more "heard"—without reporting it got in the way of getting work done.

Sources: arXiv

Why it matters: As AI assistants become fixtures in daily work, how people personalize and relate to them may be as important to design as the underlying model—one-size-fits-all chat styles could be leaving satisfaction and trust on the table.

Source: arxiv.org

Teachers Use AI to Draft Lessons, Not Replace Their Judgment

A study of eight secondary computing teachers found they don't just hand lesson planning to AI—they selectively delegate specific tasks (drafting examples, structuring activities) while keeping final judgment on classroom readiness for themselves. Using a structured AI lesson-authoring tool, researchers observed teachers treating AI drafts as a prompt to sharpen their own thinking, revising and re-specifying what they actually wanted to teach rather than simply accepting generated material.

Sources: arXiv

Why it matters: As schools adopt AI planning tools, this suggests the real value isn't speed alone—it's that reviewing AI output forces teachers to clarify fuzzy instructional goals, a dynamic that could shape how ed-tech vendors design these tools and how districts train teachers to use them.

Source: arxiv.org