Inside AT&T's Approach to Securing the Agentic Enterprise

At AT&T Secure Connections 2026, James Phillips, Vice President of Cyber Risk at AT&T, and Ofer Klein, founder and CEO of Reco, took the stage for a conversation on securing the agentic enterprise.

AT&T has gone all in on agents. Phillips described adoption that has "grown multiple X, probably 10, 20X from a year ago," with teams and individuals building agents "at speeds that we're just never used to." AT&T uses Reco's solution alongside several others, because, as Phillips put it, "it's just the speed in which everything is going is incredible."

Watch the Full Session

The Speed of Agent Adoption

As teams started building agents, the questions came quickly, Phillips said: how do you secure this, how do you think through permissioning, how do you make sure agents don't go rogue, and how do you think about intent. He joked that every three months his team looks around and asks, "is our strategy still gonna work?"

Klein put the same point in terms of velocity. "Velocity is not only how you operate, it's about how your ecosystem operates as well." He said the average enterprise already has 3.6 agents per employee, and the fastest movers have more than 90 agents per person.

Sprawl Shows Up When the Tools Turn On

Klein said it is never enough to look at Salesforce, ServiceNow or Copilot on its own, "because the combination of them is the business."

Phillips described what happened when AT&T turned some of its tools on. "We realized we had sprawl, and we had stuff that was connecting to our CRM solution that you're like, wait, we didn't realize that was happening." The next questions were how to rein it back in and what permissions those connections had, "because if you overpermission, you're gonna be able to just go run wild."

Toxic Combinations

"The agent by itself is not the target," Klein said. Looking at Fortune 500 companies, "we found, and this is Fortune 500 companies, 40% of your agents have toxic combination."

He gave two examples. An agent that should read code can also write it, and shares it externally. And a marketing ops request to make sure everybody knows the pipeline, where "she forgot to say that everybody is not externally, so everything was exposed in the internet." Klein also said that 85% of agents do not have sufficient guardrails.

Phillips shared a use case from his own organization: an HR employee building an agent to run against everyone's performance write-ups. "What's more scary to me is the permissions," he said, stressing it was a use case to talk through and that nothing was exposed.

"Go back to your organizations and really challenge, can my agents talk to the internet? And if you think they can't, I can guarantee you, you'll find some that are."

James Phillips, VP of Cyber Risk, AT&T

Blocking Doesn't Hold

Klein described a large bank that said only 50 people had access to Copilot. Those 50 people "spin off agents that spin off more agents," and the bank had tens of thousands of agents.

Phillips said restriction pushes people outside the norms: a company credit card, a personal credit card, or a home computer. "We have to figure out how to let people do it, or they're gonna figure ways around it."

Security as the Business Enabler

"The best thing we can hear from our partners and customers is, I became a business enabler," Klein said. He described security teams that blocked a lot of agents and, because they now have the tool and the governance, allow them. People are running faster, they're happier, and they're closing deals faster.

Intent, Lifecycle, and What to Do Next

Klein explained intent-based monitoring: know who or what created an agent, what it can do, what it has access to, where it needs to operate and which systems it connects to, then compare every action it takes with everything you know about it. He was clear that he is "not suggesting that this is solved."

Phillips named two takeaways for the audience. First, challenge whether your agents can talk to the internet. Second, look at your agent lifecycle: "what is my team doing on agent lifecycle, and are we retiring stuff?" He also warned that people building agents forget to turn permissions off, and said to make sure you delete those permissions.

He expects more questions from boards, outside regulators and auditors, and said the challenge is documenting how you manage agents at scale: "it's not about three agents, right? It's about hundreds of agents times every employee." And on the way forward: "this is an industry problem, it's not a one company problem."

About Reco

Reco secures the ecosystem so enterprises can enable agents with confidence. The Reco Platform discovers every agent, app, and identity across your environment, prioritizes real risk over raw alert volume, and remediates before exposure becomes a breach.