Synopsis: Before a company can govern its AI agents, it needs to know which ones are running and what they can reach. An inventory alone will not reveal the full risk. An agent’s connections to sensitive data, write permissions and external services can create dangerous combinations that look less concerning when examined separately. Effective oversight starts with understanding those relationships without shutting down the work employees are trying to accomplish.
Ofer Klein, co-founder and CEO of Reco, joins Mike Vizard to explain how security teams can catch up with business-led AI adoption. He lays out three priorities: discover agent activity, assess the exposure and take action. Discovery requires more than one vantage point. Enterprise applications may provide visibility through APIs, while personal accounts, browser activity and locally installed tools require different approaches. Knowing who uses an agent and which workflow it supports gives teams a better basis for deciding what needs attention.
For an attacker, the agent may be a way to reach valuable information rather than the ultimate target. Klein points to excessive permissions as a problem that can amplify the consequences of a compromised account or agent. Access to multiple connected tools can turn one foothold into a broader exposure. Reducing unnecessary permissions and watching for changes in behavior therefore matter alongside identifying new agents. The question is not simply whether an agent is approved, but whether its access is appropriate for the job it performs.
Klein argues that manual reviews cannot keep pace with automated activity, but he does not call for fully autonomous defense. AI-assisted monitoring needs guardrails and human oversight, particularly while agent behavior remains unpredictable. His advice is to assess the current environment, make the gaps visible to management and build a phased plan tied to business needs. Waiting for regulatory certainty leaves those gaps in place. Security teams need a workable way to reduce exposure while allowing the organization to keep using AI.
