Agent security funding surges as OpenAI doubles Anthropic's corporate users

The most telling number in today's news isn't a benchmark or a parameter count. It's the $42 million that Pluto Security just raised to protect enterprise AI agents. When investors start writing cheques that size for the security layer, it tells you the agents themselves are no longer experiments. They're in production, and somebody is worried about what they might do there.

That worry runs through several of today's stories. Snowflake's Cortex AI Gateway claims it can detect and block dangerous behavior across an agent's entire trajectory, not just individual prompts. I think that's the right framing. A single prompt can look harmless while the sequence of actions it triggers does not, and anyone who has watched a long-running agent wander off task knows how quickly that happens. A separate write-up on resolving production network incidents autonomously makes the same point from the operator's side: one safeguard is never enough, and the move from copilot to autopilot only works inside clearly defined limits. Layered controls are boring, but they're what separates a demo from something you'd let near your production network.

Meanwhile, the adoption numbers keep climbing. Samsara has opened its no-code Agent Studio to every customer and reports sevenfold growth in adoption. And a VentureBeat Intelligence survey from September found that twice as many companies use OpenAI's agent orchestration platform as Anthropic's. I'd treat that figure with some caution, since a single survey of orchestration platforms says little about depth of usage or satisfaction. But it does suggest that distribution and enterprise relationships still matter at least as much as model quality. By the way, Samsara's story is a good reminder of where much of the real value lands: not in frontier labs, but in vertical companies handing non-engineers a way to build agents for their own fleets and workflows.

Underneath all this sits the physical layer. Microsoft says it will scale its Azure Maia AI infrastructure to multi-gigawatt level from 2027, positioning its own silicon as a core part of the buildout. Gigawatts are the language of power utilities, not software companies, and the strategic motive is plain enough: reduce dependence on Nvidia and control the cost of serving all those agents. The second-order effect is that whoever secures cheap compute and energy will set the price floor for everyone building on top.

On the creative side, Higgsfield is stretching AI video from one minute to more than ten, built on GPT-6 Astra, GPT-Image-2.5 and GPT-Live. Longer coherence is a genuine step, though I'd want to see how consistent characters and scenes stay across ten minutes before calling it filmmaking.

Is the security market racing ahead of the agents it protects, or are the agents finally forcing it to grow up?