AI Agents Evade Traditional Bot Detection by Mimicking Human Behavior
Resemble AI CEO Zohaib Ahmed discusses how sophisticated AI agents, unlike simple bots, adapt their actions to achieve goals, making them difficult to detect with conventional security tools.
Traditional bot detection methods, which often rely on identifying scripted behavior, are increasingly ineffective against advanced AI agents. Zohaib Ahmed, CEO of Resemble AI, explains that these agents are designed to pursue specific goals and dynamically choose their steps, a stark contrast to bots that rigidly follow pre-defined scripts. This adaptive nature allows AI agents to mimic human actions more convincingly, posing a significant challenge for security teams.
Ahmed categorizes AI agents into three types based on their visibility: declared, undeclared, and disguised. Declared agents openly identify themselves, making them easy to spot. Undeclared agents, while not actively hiding, leave behind the typical fingerprints of automation tools, requiring specialized training data for detection. The most challenging category is disguised agents, which actively alter their identity and browser fingerprints to appear as legitimate human visitors, rendering traditional detection mechanisms largely useless.
Because most bot detection systems operate on a binary human-vs-bot classification, they frequently fail to identify undeclared agents and almost always miss disguised agents that successfully emulate human browsing patterns. This gap in detection capabilities leaves organizations vulnerable to a wide range of malicious activities that can be orchestrated by these sophisticated agents.
Identifying these human-like agents requires a shift towards more nuanced, signal-based detection. Ahmed highlights several key indicators, including the precise timing and sequencing of clicks, typing flow patterns, scrolling behavior, and subtle environmental tells such as request headers, automation flags, and Web Bot Auth signatures. Furthermore, analyzing patterns across multiple sessions for users exhibiting similar behaviors or pursuing the same goal can reveal coordinated "swarm" activities, a growing concern for security professionals.
A significant challenge arises when employees integrate AI agents into business workflows, such as email and calendar applications, using personal API keys and OAuth grants without informing security teams. These agents operate under the employee's credentials, making their activity appear as normal user behavior in most logs. To uncover these undeclared agents, Ahmed recommends starting with identity provider logs, specifically OAuth consent and app grant records, to identify third-party applications with broad access scopes approved by a single user.
Further investigation should involve auditing core SaaS tools for the creation and usage of personal API keys and tokens, as well as analyzing egress or proxy logs for unusual traffic patterns to model provider APIs or agent platforms. The key is to look for the same speed and timing anomalies observed in traditional bot detection, such as an employee reading hundreds of emails in minutes at 3 AM or accepting calendar invites instantaneously.
Ahmed advocates for organizations to embrace AI adoption securely by establishing clear policies and endorsement for employee use of AI agents, rather than attempting to ban or hunt them down. He suggests that CISOs should prioritize investments in agent detection capabilities that can effectively handle undeclared and disguised sessions. The ideal approach involves a three-step process: accurately assessing the volume of agentic traffic, defining policies to restrict access to sensitive data and risk surfaces, and implementing automated routing to balance security risks with potential revenue loss from blocking authorized agents.
Looking ahead, Ahmed predicts that metadata-based detection signals, such as IP addresses and location, will become increasingly worthless within the next six months due to their ease of spoofing and hiding. This underscores the urgent need for more advanced and adaptive AI agent detection strategies to stay ahead of evolving threats.