Deployment record
In the tracker →- Organisation
- CrowdStrike
- System
- Falcon Guardian
- Stage
- Announced
- Function
- AI Detection and Response
- 01Recognize that AI agents can behave unexpectedly in enterprise environments.
- 02Implement stronger controls as agent capabilities become more advanced.
- 03Understand that agents are designed to seek goals, potentially leading to unauthorized actions.
Show all 5Show less
- 04Be aware of prompt injection as a critical attack vector for manipulating agent behavior.
- 05Monitor for risks like credential abuse and sandbox escapes in production settings.

CrowdStrike
Follow Deployed on Google
AI agents are already behaving in ways enterprises did not plan for, according to Fabio Fratucello, Field CTO World Wide at CrowdStrike.
"Unexpected agent behavior is already occurring in real enterprise environments, and as models become more capable, enterprises require stronger controls," Fratucello told AIM Media House during an exclusive interview.
He traced the risk to how agents are built. "Agents are goal-seeking by design," he said. "They can execute exactly what was requested, encounter an obstacle, find an alternative path, and take unconventional, and potentially unauthorised actions."
Fratucello pointed to sandbox escapes documented by frontier AI labs as evidence of what increasingly capable agents can do, adding that “these scenarios extend beyond testing environments.”
In production, Fratucello said, "prompt injection remains the most critical attack pattern because it can manipulate agents into taking actions outside their intended outcome."
He listed the other common risks as "credential abuse, excessive permissions, malicious tool use, data exfiltration, and compromised agents executing malicious code."
The Endpoint as the Control Point
Fratucello said the most dangerous assumption he sees is treating a successful login as the end of the security check. "Treating authentication as the end of the security decision creates significant risk. Agents are being deployed with broad, standing permissions inherited from humans, when they require continuous control instead, he said,"
He argued that the endpoint is where that control has to sit. "The endpoint serves as the optimal control point because agents execute there; they use credentials, access data, invoke tools, and take action. Network and application controls remain important, but are limited in visibility and lack enforcement capabilities where it matters the most."
Enterprises, he said, "must observe the complete execution chain at runtime, connecting what the agent was asked to do with what it actually accomplished, and enforcing policy and technical control before that action creates impact."
CrowdStrike's Falcon Guardian, he said, "delivers the comprehensive visibility and enforcement protection enterprises need to secure AI agents during execution."

The week's banking, insurance and brokerage deployments. Tuesdays.
CrowdStrike unveiled Falcon Guardian, an AI Detection and Response solution, at its Fal.Con 2026 conference on September 1, 2026.
Falcon Guardian includes AI Agent Discovery and Inventory, which discovers known and shadow AI agents across Windows and macOS. Fratucello described shadow AI as "a massive visibility challenge," saying "organisations deploy agents faster than they can govern them."
He said unapproved agents "extend beyond unsanctioned enterprise agents to include developer and coding agents, personal productivity agents, and agents embedded within applications and workflows, all operating with credentials and permissions security teams potentially never approved."
"Security teams cannot secure what they cannot see; discovery must come first," he said.
Identity, Delegation, and Continuous Control
Fratucello said traditional identity systems were not designed for how agents operate. "Traditional identity providers force organisations to represent agents as service accounts, API keys, and workload identities. These static models fail to match agent behavior: agents operate continuously, access systems, use credentials, move data, take action on behalf of people, and delegate work to other agents."
The risk, he said, "emerges when broad permissions and standing credentials persist long after the task requiring them is completed."
CrowdStrike's Agentic IdP, he said, "establishes every agent as its own trusted, cryptographically verifiable identity, then brokers only the access needed for the required duration."
On delegation, he said, "When one agent delegates work to another, attribution follows the entire chain, binding every action back to the originating human or workload. Everything remains traceable and accountable regardless of delegation depth."
Continuous Identity, Fratucello said, "replaces one-time authorisation and standing privileges with access that is tightly scoped, continuously evaluated, and removed when no longer needed." He described it as "the shift from trusting an identity at login to continuously controlling what it can do in real time."
Before putting an autonomous agent into production, Fratucello said, enterprises need "visibility, trusted identity with no standing privileges, and runtime control at the point of execution."
He identified the move from visibility to control as the biggest gap. "Knowing an agent is deployed isn't enough, you have to be able to continuously control what it's allowed to do as it executes."
Last updated

