The Runlayer team just got back from a week at Fal.Con 2026, full of illuminating conversations on the state of security and where the industry needs to go next. As expected, the top issue on everyone’s minds: AI. In speaking with security leaders across regulated industries, digital-native enterprises, and global organizations, the same challenges and opportunities kept coming up again and again. Read on for a full breakdown, and a set of recommendations you can act on now.
The experimentation hangover is here, and it’s exposing a fundamentals problem
Over the past couple of years, board rooms and executive leadership teams have taken an “adopt fast, manage the fallout later” approach to enabling AI across their organizations. Security teams were told to accept the risk, and ensure any mitigations didn’t come at the cost of innovation. Now, the tides are turning, and the experimentation hangover has arrived.
Our Co-Founder and Chief Product Officer, Tal Peretz, moderated a panel discussion between security and technology leaders who mentioned several symptoms they’re dealing with:
- One digital-native enterprise ran an internal audit and found thousands of agents running with elevated permissions, expensive models, and no active owners.
- A large financial institution’s security team is concerned about secrets exposure, given the ease with which their large non-technical workforce can create and share credentials via AI clients.
- Another technology leader cited the near-impossible challenge of relying on humans alone to monitor millions of agent events per day.
Throughout the discussion, the panelists agreed: AI is exposing the lack, and criticality, of security fundamentals at scale.
Applying the fundamentals to agents that move at machine speed
The good news is that these security leaders have already found ways to incorporate the fundamentals in a way that can match AI’s pace.
1. Enforce standards at the infrastructure layer. Defense-in-depth may look different in an AI environment at face value, but at the core, the principles remain the same. Route all AI traffic through a single managed pipeline that can handle every client, harness, MCP, and skill file. This ensures that every prompt is logged, every model is controlled centrally, and agent intent is categorized. One enterprise with tens of thousands of developers runs this with no exceptions, and perhaps equally importantly, no disruptions to the developers’ work.
2. Credential management and least privilege access are paramount. As agents take on more complex work, they must be treated as independent identities, with their permissions and access secured accordingly. Create scoped allowlists of the exact permissions and tool calls needed to allow the agent to complete its task, and block everything else by default. Delineate which agents will inherit a specific user’s permission, and which agents will run on their own service account credentials, with the former being used only when human interactivity is required in an agent’s lifecycle. Ensure every human and agent action is logged and correctly attributed, such that you can conduct regular access reviews to descope elevated permissions over time.
3. Continuous monitoring must incorporate agents. One digital-native enterprise sees 200,000 events in a single agent session facilitated by a developer. Multiply that across thousands of developers and agents running every day, and there’s a volume of data that no human could hope to manage by themselves. The panel's consensus on how to attack this issue: use agents to monitor agents, with human-in-the-loop at the appropriate step. Incorporate machine learning rules first to categorize the entire dataset, then use agent-based validation to filter the signal through the noise, and incorporate human judgment on escalations.
Interestingly, none of the panelists recommended outright blocking AI use. They recognized that if they don’t provide their workforce with an opinionated, sanctioned way to build agents, teams will just build their own.
Make the safe path the golden path
Security teams need to find a way to say “yes” to AI transformation without compromising on unacceptable levels of risk. By making the safe path the golden path, you take an active role in enabling sanctioned and secure AI use across the workforce.
This is exactly what Runlayer’s AI control plane is built to do. With an employee-facing catalog of approved AI agents, MCPs, and skills, the easy path becomes centrally managed. And behind the scenes, you get continuous shadow AI discovery, an MCP & LLM gateway, runtime security, and fine-grained AI governance to ensure every agent runs on the golden path.
See why industry leaders like Gusto, PagerDuty, and dbt Labs trust Runlayer to help them go all in on AI → book a custom demo.
