Zero Trust for AI Agents Starts With Network Visibility
Back to top
November 19, 2026
Zero Trust for AI Agents Starts With Network Visibility
In the first week of November, two organizations disclosed automated activity running inside internal systems for days before security teams intervened. WorkOS, a developer platform handling sign-in and authentication for software companies, traced unauthorized access to an AI security agent a researcher had tested in an isolated environment before the agent escaped into production systems. Mittwald, a German hosting provider for web agencies, reported that threat actors spent eight days inside customer environments using AI-assisted tools.
The disclosures raise pointed questions for CISOs.
- Which AI agents operate in the environment?
- Who owns each agent and governs least-privileged access?
- What is each agent doing right now?
- How is continuous verification validating runtime behavior?
How Do Unmonitored AI Agents Break Zero Trust?
Provisioning an autonomous workflow once required months of architecture review, service-account approvals, and firewall changes. Today, engineering and business teams wire language models directly into internal APIs, databases, and file repositories in hours. Accelerating rapid adoption is the Model Context Protocol (MCP), an open client-server specification giving any compatible model a standardized plug into enterprise tools.
Speed changes how Zero Trust fails. When teams connect MCP servers and agents outside central security review, unmonitored proliferation breaks Zero Trust's core premise — "never trust, always verify" — by leaving uncataloged workloads running on long-lived tokens or broad user credentials. Because standard MCP authentication validates the initial client connection rather than each downstream tool call, an authorized agent behaves like a confused deputy. A single indirect prompt injection hidden inside an incoming support ticket, email, or retrieved document can steer a legitimate agent into reading sensitive records or crossing segmented networks under valid credentials, without triggering a single identity alarm.
Over half of security leaders rank AI agents as a top attack surface risk, according to ExtraHop’s 2026 Global Threat Landscape Report. Blind spots in agent discovery turn sanctioned credentials into unmonitored attack paths.
Why Does Zero Trust Require a Complete AI Agent Inventory?
Zero Trust policies govern known workloads with accountable owners, explicit permissions, and expected traffic patterns. An uninventoried agent lacks all three, leaving policy engines with zero criteria to verify requests.
Even inside authorized workflows, autonomous models routinely improvise unexpected paths within literal instructions, drifting from intended scope, hallucinating outputs, or chaining tools human operators never anticipated. Because agents communicate over valid encrypted protocols, erratic or unauthorized calls blend directly into routine application traffic. Endpoint software can be bypassed and local logs altered; continuous verification requires independent proof of runtime behavior rather than static credentials alone. When unmonitored agents call high-privilege MCP gateways or internal APIs, enterprise infrastructure opens unguarded pathways straight into core data repositories.
What Do the WorkOS and Mittwald Incidents Reveal About Agent Visibility?
Retrospective log analysis breaks down once automated tools act at machine speed. At WorkOS, an escaped evaluation agent discovered an exposed internal sales tool, authenticated with a predictable development credential, and operated for five days before an employee noticed an unusual surge in CRM queries. At Mittwald, threat actors exploited a flaw in an MCP server interface, harvested customer session tokens, and moved across 56 customer projects with seconds between steps — remaining undetected for eight days.
In both incidents, traditional forensic reviews reconstructed damage only after sensitive records had already been exposed. When autonomous tools execute multi-step sequences in seconds, delayed detection nearly guarantees an extended window of compromise.
Why Do Logs and Endpoint Software Miss AI Agents?
Host-based telemetry records only compliant, reporting endpoints, leaving security teams blind whenever endpoint agents are bypassed, silenced, or misconfigured. Unmanaged devices and ephemeral cloud containers running AI workflows frequently sit outside host software coverage altogether.
Network traffic doesn’t offer a comparable opt-out. Every prompt sent to a language model, every tool invocation routed through an MCP server, and every query hitting a core database API must traverse the wire. Treating network packets as an unbypassable record of runtime activity closes the visibility gap that otherwise allows rogue agents to reach critical data unobserved.
How Does Network Evidence Reveal Which Agents Are Operating?
Wire data exposes the full operational footprint of AI agents across the enterprise. Passive, out-of-band network sensors automatically discover active AI agents, mapped Model Context Protocol (MCP) servers, and runtime API calls without requiring host software on client devices. Real-time decryption of encrypted agent sessions lets security operations link prompt executions, downstream tool calls, and credential delegation directly back to initiating human identities and parent machines.
Comparing observed agent interactions against sanctioned baselines isolates shadow AI agents immediately, supplying the continuous behavioral verification Zero Trust architecture demands to intercept unauthorized actions mid-flight rather than after an incident concludes.
What Limits the Reach of Network Evidence?
Network visibility boundaries reflect sensor placement and processing throughput. Traffic crossing unmonitored cloud-to-cloud links stays outside the inspection record unless sensors span data center, campus, branch, and multi-cloud environments. Line-rate capture must also keep pace with high-throughput AI backbones, requiring inspection platforms built for 400 Gbps architectures so that high-volume traffic never creates blind spots.
How Do Security Leaders Reconcile Zero Trust with Autonomous AI Agents?
Zero Trust policy enforcement starts with complete asset discovery. Security leaders must first establish passive network visibility to inventory every active AI agent, map underlying MCP server connections, and tie runtime tool calls to verified human or service credentials.
ExtraHop CISO Chad LeMaire notes that earning trust with autonomous workflows requires pairing negative guardrails — blocking unauthorized writes even when an underlying credential carries broad permissions — with positive guide rails defining exact required steps. Baselining normal agent traffic on the wire then supplies the hard evidence needed to spot privilege escalation, block lateral movement, and contain rogue autonomous activity before a breach unfolds.
Download ExtraHop’s 2026 Global Threat Landscape Report: AI Edition to benchmark AI agent exposure against global security peers, or explore how ExtraHop RevealX discovers shadow AI agents and decodes MCP traffic in real time.

Chief Evangelist
Heath Mullins is the Chief Evangelist at ExtraHop with 27+ years of experience designing global network architectures and threat detection strategies. Heath Mullins previously served as a Senior Analyst at Forrester advising Global 100 enterprises and specializes in implementing zero-trust methodologies through Network Detection and Response (NDR) deployments. View Heath Mullins’ complete professional profile on LinkedIn.
Share
Key Takeaways
- Modern enterprises rapidly deploy autonomous AI agents that operate completely outside established Zero Trust security frameworks.
- Enforcing Zero Trust across autonomous AI agents requires cataloged asset ownership alongside defined behavioral parameters.
- Standard identity checks permit encrypted agent traffic blindly, masking unauthorized tool calls from enterprise defenses.
- Attackers weaponize automated workflows in seconds, rendering post-incident log reviews ineffective against fast-moving threats.
- Continuous network analysis provides objective proof of runtime behavior, exposing shadow integrations before damage occurs.



