The conversation around AI agents has evolved dramatically in the past year. Early excitement centered on rapid deployment and promised productivity gains, with teams racing to plug largeโlanguage models into workflows. However, a series of highโprofile breachesโmost notably the intrusion uncovered during an OpenAI agent evaluation at Hugging Faceโhas forced a sober reassessment. Security leaders now ask not just what an agent can do, but what it can reach once it is running and whether any anomaly would surface before damage occurs. Jumping straight into firewalls, detection rules, or access controls without first understanding the landscape is akin to locking doors while leaving windows wide open. The prerequisite for any meaningful Zero Trust stance is a clear, continuously updated picture of every agent in the environment, its data touches, and its behavioral baseline. Without that foundation, enforcement becomes guesswork, and attackers can exploit the blind spots that speedโfirst adoption inevitably creates. Market analysts note that enterprise AI agent spend is projected to exceed $15โฏbillion by 2028, yet less than 30โฏ% of organizations report having a centralized inventory of those agents, highlighting a critical gap between investment and oversight.
Recent research from Veeam underscores the scale of the visibility problem. Seventy percent of surveyed enterprises admit that AI workflows are already interacting with sensitive corporate data without full oversight, while twoโthirds say IT cannot fully track the autonomous workflows employees are building on their own. This phenomenonโoften labeled Shadow AIโmirrors the historic Shadow IT challenge but moves at a far faster pace because agents can be spun up in seconds via lowโcode platforms or browser extensions. When an agent operates unseen, it can exfiltrate credentials, manipulate data pipelines, or generate costly API calls without triggering any alert. The financial impact can be staggering: a single compromised agent leveraging an unrestricted API key burned through the equivalent of $600โฏk in tokens over three weeks at a nonprofit research group. Such incidents reveal that the cost of invisibility is not merely theoretical; it translates directly into wasted compute, potential regulatory fines, and erosion of trust in AI initiatives.
Zero Trust offers a robust framework for governing AI agents, but only when applied in the correct sequence. The SANS cheat sheet for Zero Trust for AI Agents places inventory at the very top, embodying the axiom โyou cannot govern what you cannot see.โ If an organization jumps to policy enforcement points, authorization schemes, or network proxies before establishing a named owner, defined scope, and inventory entry for each agent, those controls have nothing substantive to enforce against. A proxy sitting in front of an unknown population of agents is analogous to a security guard posted at a building with no visitor listโhe can challenge anyone, but he has no way to know who belongs inside. This misordering is a common failure mode that leaves security teams investing in sophisticated detection engines while the underlying asset base remains a mystery, rendering alerts noisy and ineffective.
Think Red: From an attackerโs standpoint, an invisible agent is a lowโeffort foothold. Consider the case at METR, the nonprofit that assisted in evaluating the Hugging Face incident. An adversary discovered an employeeโs personal EC2 instance running a vibeโcoded agentic application, bypassed its weak authentication, and prompted the agent to relinquish its modelโprovider API key. Because the agent had no spending limits, the intruder consumed roughly $600โฏ000 worth of tokens over three weeks. METRโs internal dashboard omitted rateโlimited request data, and raw token volume alone did not cross any threshold, allowing the abuse to persist unnoticed. The attacker needed no elaborate network breach; they simply leveraged the visibility gap to turn a benignโlooking tool into a profitโdraining weapon.
Act Blue: Defenders can borrow cloudโvisibility practices that have matured over the past decade. Just as finance teams monitor AWS/Azure spend and shut down idle virtual machines, organizations should treat AIโrelated expenditures and APIโkey issuance as discovery signals. Publishing an approvedโprovider pathway before issuing blanket blocks gives legitimate users a sanctioned route to follow, reducing friction while still enabling control. The guiding principle becomes โknow first, then restrictโ: first establish what agents exist, who owns them, and what data they may touch, then apply granular policies. Involving procurement and finance early ensures that shadow usage surfaces through expense reports and contract reviews, creating a complementary vantage point to technical telemetry.
Even after committing to discovery, no single data source can capture the full agent ecosystem. Agents reside in disparate locations: traditional network zones, endpoint devices, browser extensions, and embedded SaaS services. Traffic to model providers is typically TLSโencrypted, so an inline sensor sees only a destination IP and byte count, not the actual prompt, tool call, or data exfiltration attempt. Because many legitimate SaaS applications share the same domains (e.g., api.openai.com), networkโlevel analysis cannot distinguish malicious from benign traffic. Endpoint detection tools miss browserโresident agents, and SaaSโembedded logic remains invisible to both network and host monitors, leaving significant blind spots if reliance is placed on a single lens.
Act Blue: Recovering visibility requires a layered, correlationโcentric strategy. Metadata such as DNS queries, SNI values, and JA4 TLS fingerprints can hint at communication with known model providers even when payloads are encrypted. Egressโproxy logs provide connection timestamps and volume data that, when aggregated, reveal unusual patterns. Complement this with endpoint telemetry: processes spawning unfamiliar runtimes, API keys residing in environment variables, or local agent frameworks. Browserโlevel visibility is equally criticalโmonitor extensions, inโpage copilots, and enterpriseโbrowser logs for signs of unauthorized AI usage. Finally, tap into identity and SaaS signals: OAuth grants, APIโkey creation events in provider admin consoles, and serviceโaccount activity. By fusing these disparate streams into a unified inventory, security teams can construct a comprehensive, continuously updated map of the agent landscape.
Point solutions such as LLM gateways (e.g., LiteLLM) can centralize visibility and governance, but only for agents that are already configured to point through them. A gateway becomes an effective policy enforcement point once an inventory exists, yet it does not discover the agents that remain unconfigured or bypass the gateway entirely. This creates a chickenโandโegg problem: you need visibility to feed the gateway, but the gateway alone cannot generate that visibility. Relying solely on a gateway without a preceding discovery phase leaves unknown agents free to operate in parallel, potentially undermining any controls you believe you have installed. Therefore, gateways should be viewed as a laterโstage enforcement layer, not the initial source of truth.
Traditional periodic audits are illโsuited for the ephemeral nature of AI agents. An annual review can only reflect assets observed a year earlier, which is meaningless when agents can be cloned, redeployed, or decommissioned in seconds. By the time a review cycle concludes, the inventory it produced is already obsolete, giving attackers a wide window to exploit shortโlived instances. Continuous monitoring is the obvious remedy, yet removing human oversight entirely introduces accountability risks. Automated systems watching other automated systems can produce blind spots if no person is ultimately responsible for interpreting alerts, validating findings, and driving remediation.
Think Red: An attacker who knows an organization audits quarterly could instruct a compromised agent to spawn shortโlived clones that inherit the parentโs privileges, perform a malicious taskโsuch as exfiltrating a dataset or launching a credentialโphishing campaignโand then terminate before the next review window. Each clone lives just long enough to achieve its objective, leaving no trace in the periodic snapshot. This โflashโmobโ tactic exploits the gap between assessment intervals and the rapid lifecycle of modern AI workloads.
Act Blue: A resilient visibility program blends human insight with automated observation. Have security analysts periodically review agent telemetry while also enabling agents to log and share metadata about their own behavior (e.g., tool invocations, data reads/writes). This mutual observation reduces reliance on any single point of failure. Crucially, auditability must retain a named ownerโsomeone who is accountable for the accuracy of the inventory and the timeliness of responses. To mitigate risk, institute quality gates: require that any new agent pass a baseline inventory check, receive a justified business purpose, and be logged in a centralized asset register before it is granted access to production data or services.
Legislative momentum is reinforcing the technical argument for visibility. Recent US discussions around an emergency โkill switchโ for autonomous AI acknowledge that the ability to shut down a rogue agent is only meaningful if you know exactly what to switch off. Without a precise inventory, a kill switch risks either disabling legitimate services or missing the malicious actor entirely. This policy push mirrors the broader trend toward AIโspecific regulations that demand transparency, traceability, and humanโinโtheโloop oversight, especially for systems handling sensitive data or critical infrastructure.
Agent identity is the linchpin that transforms raw visibility into actionable security controls. Rather than treating an agent as a mere extension of the user who launched it, model each agent as a distinct identity with its own policy envelope. Bind permissions to the active task at runtime, enforce strict egress constraints on what data may leave the environment, and place an authorization layer between the model and any connected services. Logging must evolve from capturing raw prompts to recording the full sequence of tool calls, API invocations, and state changes performed by the agent. When every action is attributable to a specific agent identity, detection thresholds become meaningful, and anomalous behavior can be correlated with a known entity rather than lost in noise.
Where visibility sits in your order of operations determines the success of every downstream governance step. Each challenge describedโshadow AI, distributed deployment, inadequate point solutions, stale audits, and missing identityโstems from the same root: attempting to enforce controls before a reliable inventory exists. The prescribed path is clear: begin with comprehensive discovery by correlating network, endpoint, browser, identity, and SaaS signals; maintain continuous monitoring that couples automated collection with human oversight; assign each agent a unique, leastโprivilege identity; and only then layer on policy enforcement, detection, and response mechanisms. The full Zero Trust for AI Agents security checklist walks through these three maturity tiersโinventory and governance, architecture and enforcement, and detection and responseโin the sequence they must occur. For practitioners seeking handsโon guidance, the upcoming SANS course SEC530: Defensible Security Architecture and Engineering, Implementing Zero Trust for the Hybrid Enterprise, offered this December at the Cyber Defense Initiative 2026, provides a deep dive into these controls and practical lab exercises.