Google Says AI Agents Stole Credentials In Six Hours

Google Threat Intelligence Group published a report this week documenting a financially motivated group that ran a multi-agent AI framework to compromise thousands of credentials in under six hours, assigning separate agents to reconnaissance, credential validation and lateral movement at once. Kai Tanner's desk has the number that should worry a compliance officer more than a security one: six hours sits inside most banks' detection to containment window, not outside it. The runbooks built for human paced attackers are already obsolete against the machines now running the other side.

The shift from tool to actor is not confined to offense. Aya Nakamura's desk reports that Nvidia is acquiring Hugging Face, the repository where nearly every AI lab distributes its model weights, for $13 billion, meaning the company that sells the chips training the world's models now also owns the shelf they sit on before anyone downloads them. Google's own parallel report tracked state linked operators in China, Iran and Russia making the same move Tanner's filing describes, off single turn prompts and onto autonomous systems that chain actions without a human approving each step. Two industries, hardware and cybercrime, arrived at the same architecture in the same week. That is what happens when the underlying technology gets cheap enough to run unsupervised.

Elsewhere, Mei Chen's desk has Taiwan standing up an expanded, US backed cyber unit built to hunt hackers rather than patch breaches after the fact, a choice that reads as Taipei bracing for disruption and sabotage on a timeline measured in months, not for an invasion. Vincent Lai's ledger has the People's Bank of China issuing a rare direct warning against one sided yuan bets, the kind of statement that has historically preceded intervention. In science, Magnus Honeyfield's desk covers a solid state battery that beats the energy density of the lithium ion cells that have capped EV range for a decade; mass production has not been scheduled and no automaker has named a launch date.

Google's finding is that agentic AI is already running offensive operations, not sitting as a future attack surface to secure, and Nvidia's purchase is a bet that whoever owns model distribution controls the supply chain underneath every one of those agents. A bank auditing its model sourcing policy and a SOC team auditing its detection curve are, this week, auditing the same problem from opposite ends.