AI Briefing: Agent Security, Nvidia Probe, DeepSeek Surge
AI agents are creating new risks and momentum: a five-minute Vault breach, a CEO probe, and DeepSeek's surge.
Why developers should treat leaked secrets as already compromised
What the Nvidia probe and DeepSeek's Harness tell us
Today's briefing reads like a stress test for the AI industry. Nvidia shares fell 5% after the board opened a probe into whether CEO Jensen Huang was targeted by foreign intelligence, following his admission of an extramarital relationship. At the same time, Microsoft revealed an AI agent that went from scanning public GitHub to copying data out of an Azure Key Vault in five minutes. These stories are connected: AI's reach is growing faster than our ability to secure it.
## Why developers should treat leaked secrets as already compromised
The Azure incident is the most actionable story here. Attackers used an agentic framework called ARTEX to find leaked credentials, application logins, and AI API keys in public code, then used them to sign in across multiple Azure tenants. The alarming part: deleting a secret from a repository does not revoke it. If a credential ever appears in public code, assume it's compromised and rotate it immediately. For developers, this is a reminder to scope identities tightly and monitor sign-in anomalies — an agent can move in minutes what a human team might never notice.
## What the Nvidia probe and DeepSeek's Harness tell us
The Nvidia story is less about code and more about the geopolitical and supply-chain risks embedded in AI hardware. The board says it has no reason to believe national security or Nvidia secrets were compromised, and Huang remains CEO, but the episode underscores how personal and political factors can affect major AI companies. On the product side, DeepSeek reports its open-source agent app Harness reached 1.54 million daily active users and roughly 92 trillion tokens per day. That scale shows agentic interfaces are becoming a real platform. Still, the same briefing reports that the best coding agent fully cleared only 26.4% of real tasks, so autonomy remains limited.
The takeaway is to use AI tools with your eyes open. Treat public code as hostile, enforce least-privilege access, and keep a human in the loop. Today's news also includes a new requirement that Claude-driven machines must stop when the human link drops — a sensible default for any automation. AI agents can save hours, but they can also cause a five-minute breach. Design your systems as if an agent will eventually run against them.
The five-minute Azure breach is the canary in the coal mine for AI agent security.