The benign true positive: When trusted AI agents look like attackers
CSO Online ·

A coding agent opens hundreds of files, launches command-line tools and connects to an internal service to complete a task. A support agent authenticates to several systems, retrieves customer records and updates another application. A workflow agent enumerates cloud resources in seconds, then moves to the next stage of its assignment. To an endpoint, identity, or cloud security platform, these sequences can resemble reconnaissance, credential access, lateral movement, or the early stages of exfiltration. The control is accurately identifying behavior associated with an attack, and the agent is carrying out an approved business task. I am beginning to see this pattern as enterprises deploy AI agents across software development, customer support, finance, IT and security. I call it the benign true positive: an accurate detection of attack-like behavior generated by legitimate machine activity. For CISOs, this exposes a foundational assumption beneath behavioral security . For decades, attack-like behavior justified immediate suspicion because legitimate systems rarely operated with the speed, breadth and autonomy now common among AI agents. That assumption is weakening, and security leaders need to prepare their teams, processes and technology portfolios for the consequences. Behavioral baselines were built around people Modern detection systems rely heavily on patterns associated with expected human activity. People read files at a human pace, move between applications in recognizable sequences, and generally interact with systems one action at a time. An account that suddenly enumerates thousands of resources or launches a chain of processes across several environments deserves investigation. AI agents follow a different operating pattern. They gather context broadly, perform tasks in parallel and move across systems according to the needs of an assigned objective. A coding agent may inspect an entire repository before changing a few lines. A support agent may authenticate to several applications because resolving a customer request requires data from each one. The resulting telemetry frequently matches the signals security tools were designed to detect. Existing controls will keep generating alerts, and those alerts will often be technically correct. Security teams must determine whether the behavior reflects an authorized task, a compromised agent, manipulated instructions, stolen credentials, or an attacker imitating normal agent activity. A broad suppression rule may reduce noise but also create a blind spot. Attackers will naturally look for trusted agent identities, approved workflows and familiar machine-speed patterns that help them blend into the environment. Intent becomes the missing layer Security tools observe actions such as process launches, file reads, authentication events, API calls, or network connections. Intent sits in the surrounding context. Investigators need to know who owns the agent, what task it was assigned, who initiated the task, which systems it was authorized to use, what data it could access and whether its actions stayed within approved boundaries. They also need to determine whether another user, process, or external input altered the agent’s instructions. Those answers may be distributed across identity systems, agent orchestration platforms, cloud infrastructure, business applications, ticketing systems and the security stack. Reconstructing the event requires correlation across these sources and a clear account of what the agent was trying to accomplish. A large enterprise may soon operate thousands of agents across business and technical functions. Human analysts cannot manually reconstruct every event at that scale. Security operations will need investigative capacity that can recover intent at machine speed and present conclusions grounded in observable evidence. 3 investment priorities for CISOs 1. Agent identity and governance The first priority is agent identity and governance . Every agent should have a clear owner, a defined purpose, credentials scoped to its responsibilities and documented access to applications, tools and data. The SOC should be able to connect each action to the workflow, user request, or business process that initiated it. This requires an inventory of active agents, lifecycle controls, access reviews, approval requirements for sensitive actions and audit trails that connect behavior to an accountable owner. Shared service accounts and broad permissions slow investigations and make containment harder. Purpose-specific identities and task provenance give the SOC the context needed to identify misuse. 2. Investigation architecture The second priority is investigation architecture. Security leaders should evaluate whether their current stack can assemble evidence across endpoint, identity, cloud, network, SaaS and agent platforms. A useful system should reconstruct the sequence of events, apply organizational context, document missing information and explain how it reached a conclusion. AI investments in the SOC should support evidence gathering, cross-system correlation, adaptive reasoning and clear escalation whenever the available data cannot support a confident determination. 3. Team design The third priority is team design. Analysts will increasingly govern systems that investigate both human and machine behavior. Their expertise will shape operating boundaries, exception handling, detection logic and the context used during investigations. Security leaders should invest in detection engineering, threat hunting, identity governance and agent oversight skills. Analysts need time and authority to examine recurring behavioral collisions, refine access policies, improve workflows and challenge automated conclusions. Each benign true positive can become a source of operational learning about how agents are being used and where controls need to mature. A new operating model for the SOC The traditional SOC sends alerts to analysts, who gather evidence and decide what happened. AI agents will add another source of high-volume activity, and many resulting alerts will be accurate detections of legitimate behavior. The SOC needs an operating model built around rapid interpretation. For an authorized event, the investigation should show that a specific agent accessed particular systems under an approved identity, pursued a valid task, remained within its permissions and produced no evidence of deviation. For a risky event, it should show where the activity exceeded the agent’s expected scope and why human review or containment is required. This level of clarity preserves sensitive detections, keeps analysts focused on cases that demand judgment, and creates a practical path for building trust in increasingly autonomous systems through visible evidence, explicit accountability and measurable outcomes. Preparing for the next source of alert volume CISOs should begin by mapping where agents already operate across the organization. Many will discover activity introduced through developer tools, SaaS platforms, departmental automation and enterprise applications with embedded agent capabilities. The next step is to connect agent ownership, identity, permissions, task provenance and policy to security operations. Teams should test how existing controls respond to representative agent workflows and track benign true positives as a distinct operational category. AI agents are changing the meaning of normal behavior. Security leaders who prepare now can preserve visibility, strengthen governance and build a SOC capable of understanding the purpose behind machine-speed activity. Trusted machines will increasingly look like attackers. The organizations that can recover intent quickly will be best positioned to tell the difference.