CoreWeave’s Phil Gurbacki on AI Agents, Guardrails, and the Future of AI Development

For quality assurance (QA) engineers, the AI era represents an opportunity that has been a long time coming.

The idea behind autonomous AI sounds quite attractive and simple. Give a powerful model a goal, access to the right systems, and enough compute, and just let it decide how to get the job done on its own. That might sell well during a product demonstration. But it’s a risky and expensive way to run an enterprise.

After years sitting across the table from enterprise security leaders, I've stopped hearing the question I used to hear constantly: does AI belong in security operations? They now ask how to deploy it without breaking what already works.

A look at why the fight over Chinese open models is really a fight about who captures the value of AI -- and why banning them is the worst possible answer.

Like cloud and DevOps before it, 'forward-deployed engineering' risks becoming a label more than a practice.

86% of dealmakers already use GenAI in M&A—the open question now is build it in-house or buy it off the shelf.

Up to 90% of enterprise data is unstructured, and most of it stays invisible to the AI tools built to use it.

Faster code generation just means longer review queues—output capacity, not review capacity, is what's growing.

What looks like one global system on paper is often quietly fragmenting into regional copies underneath.

IFM's Liu argues language models alone can't understand the physical world—AI needs world models too.

Sage's Thiara says finance leaders won't trust an AI agent unless they can see exactly how it reached an answer.

A single AI-flagged alert can quietly trigger a chain of billable actions most SMBs never see coming.

Asking engineers to document after the fact has failed for thirty years—the people who know most write the least.

WAN traffic could surge 700% by 2034, and most networks were never built for AI's bursty, latency-hungry patterns.

AI coding tools don't fix broken teams—DORA data shows they simply amplify whatever dysfunction already exists.

Most enterprise AI tools show up through employee sign-ups, not IT—echoing shadow SaaS, but faster.

RSA's Taylor says as authentication gets stronger, attackers are pivoting straight to the help desk instead.

Cloud GPUs are cheap to rent for prototyping, but real-time inference often runs cheaper and faster at the edge.

Waiting for a full ERP overhaul before deploying AI means leaving quick, high-value wins on the table.

Restoring from backup after ransomware can reintroduce the very malware that took you down.

One team admitted throwing away 499 out of every 500 annotations it paid humans to create.

Five prompts and five minutes: that's all it now takes AI to write a phishing email that once took 16 hours.

A burned-out employee pasting sensitive data into an AI tool is scarier than any hacker on the outside.

With just 4,255 ABAP developers left in the U.S., 23,000+ SAP shops face a 2030 deadline with no one to hire.

AI-assisted attackers now go from breach to data theft in 25 minutes—versus two days for humans.

AWS's Schreiner says the hardest part of AI transformation isn't the technology—it's trust and change management.

Every AI project reinventing its own networking and compliance review is a sign platform teams are about to burn out.

Sumo Logic's Clawson says enterprises are underestimating how fast AI agents will remake security operations.

Celonis's Brown argues storytelling and empathy, not just dashboards, are what get employees to trust automation.

No matter how advanced the model, AI is still only as reliable as the data and systems underneath it.

Fraud tactics are crowdsourced in encrypted chats long before static intelligence feeds ever catch them.

The safer a model gets, the more it refuses to help test its own defenses—a paradox safety teams must solve.

Feeding every log into your SIEM —just in case—quietly inflates budgets without improving detection.

90% of breaches now start at cloud-based web app servers—detection-first security can't keep up.

One casual query to an AI assistant can surface data that was never meant to be seen.

Google's Willow chip ran a task in five minutes that would take a supercomputer 10 septillion years.

The agentic AI market is set to jump from $7.3 billion to $41.3 billion by 2030—fast enough to outrun most strategies.

With cloud spend headed past $723 billion, more enterprises are pulling data back to where it's regulated.

A polished demo can still turn out to be rigid automation dressed up as AI once it hits production.

Broadcom just cut VMware's reseller base from 4,000 to 1,500—and it's reviving competition, not killing it.

Just as ISO 27001 brought order to security, this new standard aims to do the same for AI governance.

Legacy platforms and agentic AI seem worlds apart—but agents are already modernizing old systems from within.

Coding AI now solves 70% of hard tasks—a tenfold jump in just two years, per Stanford's latest data.

Amazon, Microsoft, Meta and Alphabet are betting big on AI—that spending pattern reveals where things are headed.

Meta calls Llama open source without an OSI license—a labeling fight that's still far from settled.

Komprise's Subramanian says the real AI differentiator isn't compute power—it's how well you manage your data.

By 2027, Gartner expects task-specific small models to outnumber general LLMs three to one in enterprises.

Extreme Networks' Bukhari expects AI to make human judgment and empathy more valuable, not less.

Even OpenAI, once going its own way, ended up adopting Anthropic's open MCP standard.

Cloudera's Rajaji argues open models win long-term because switching costs, not raw capability, decide loyalty.

78% of users say chatbots can't handle complex issues—and too many block the path to a human.

Snowflake's Kleinerman makes the case for staying model-agnostic instead of locking into one AI vendor.

Analyst David Linthicum says enterprises are outgrowing big cloud platforms in favor of AI-optimized providers.