Satya Nadella Says AI Agents Should Be Treated Like Insider Risks

Morning, {{first name | folks}}! A lot happened in AI this week. Microsoft raised new questions around AI agent security, NVIDIA highlighted AI advances in breast cancer care, Amazon reportedly explored a $7 billion deal for Decart, computing power remained a major challenge for AI startups, and Moonshot faced a shareholder dispute ahead of its planned IPO.

Microsoft CEO Satya Nadella says companies need to treat powerful AI models as potential insider risks, even when there’s no malicious intent. As AI agents gain access to sensitive data and the ability to take action, their behavior can be difficult to trace back to a specific cause.

His argument is that companies shouldn’t rely on model providers to guarantee safe behavior. They need independent controls over what AI can access and do, clear records of its actions, regular testing and a way to stop agents mid-task. The key idea is to give AI the ability to act without giving it control over its own permissions.

NVIDIA is highlighting startups using AI across breast cancer care, from screening to treatment planning. iSono Health has developed an automated 3D ultrasound system that captures scans in about two minutes per breast, while Whiterabbit.ai uses AI to assess breast density and estimate cancer risk.

Others are focusing on what happens after detection. Ataraxis AI and SimBioSys are developing tools to help predict treatment response and visualize tumors, bringing AI further into clinical decision-making. The goal is to give doctors more information to help guide treatment, not just identify cancer earlier.

Amazon is reportedly in talks to acquire AI startup Decart for around $7 billion, according to The Wall Street Journal. The startup develops AI technology for real-time video generation and simulated environments that can help train and test robotics systems.

The potential deal would be a major step up from Decart’s roughly $4 billion valuation in May. It would also give Amazon access to technology that could support its broader AI ambitions, from generating video to building and running advanced AI models more efficiently. The discussions are still ongoing, and there is no confirmed agreement yet.

The race to build AI is creating a scramble for the computing power behind it. As demand for advanced chips and data center capacity grows, startups are finding that funding alone doesn’t guarantee access to the resources needed to train and run their models.

Securing enough compute is becoming a business challenge in its own right. Companies must plan ahead to avoid being left short while balancing the cost of reserving capacity they may not use. For AI startups, access to computing power could increasingly determine how quickly they can grow and compete.

Chinese AI startup Moonshot, the company behind Kimi, is in talks to settle a shareholder dispute ahead of a planned Hong Kong IPO, according to Reuters. Shareholder Recurrent.ai and some investors have accused Moonshot’s founders of misleading investors and diverting company resources. Moonshot has declined to comment, while its CEO has previously denied the allegations.

Recurrent is seeking compensation that could exceed $700 million, based on Moonshot’s reported $30 billion valuation. The talks come as the startup prepares to go public, adding pressure to resolve the dispute before its IPO.