Your AI Agents Are Getting Phished, And Waiting To Respond Won't Save Them
IRONSCALES founder & CEO Eyal Benishti is a pioneering security software engineer & executive leader with 15+ years of industry experience.
gettyFor decades, phishing has had one target: people. Every line of defense we’ve constructed, from filters to awareness training, assumed a human sat at the end of the chain. For the attack to work, someone had to click the link, open the attachment, provide the credentials or approve the wire transfer. That window between delivery and action is where detection did its work. Models that learn how people normally communicate could spot what looked off and stop the threat before anyone acted.
Inboxes have a new reader: AI agents. And these agents are being given real autonomy, including the ability to summarize messages, file tickets, move documents and even trigger workflows on employees’ behalf. They work quickly, they don’t second-guess themselves, and attackers have started to take notice.
In such an environment, traditional detection and response approaches to defense begin to break down. When attacks can succeed without any human intervention at all, it’s imperative that organizations adapt to a more preemptive approach to cybersecurity.
The scale of this shift is hard to overstate. In a survey of 501 U.S. IT and security leaders from Censuswide and commissioned by my company, 96% said their organizations use AI agents or tools that can take actions on behalf of employees. And more than half (52%) said that use is widespread across the organization.
These agents aren’t confined to performing innocuous tasks. Many can access and interact with ticketing systems, HR services, cloud storage, internal files, security tools and, in some cases, payment systems.
And attackers are exploiting that access. In the same survey, 77% of respondents said that in the past year, an AI agent at their organization acted on instructions or content that turned out to be malicious, misleading or unauthorized. This isn’t some threat looming beyond the horizon. It’s already here, and already making its presence felt.
The most common way these agents are being exploited today is a familiar one. Among organizations that had an incident, the top type was an AI tool summarizing or acting on a malicious email or message. Phishing hasn’t gone anywhere; it’s just found a new target. A malicious message no longer has to fool an employee. It only has to fool the agent sitting between that employee and their inbox.
Most security teams already automate what they trust to run on its own. Detection flags the message, remediation pulls it, and an analyst reviews the calls that need judgment. Teams set those thresholds deliberately, because someone has to answer for every automated action. That design works when the attacker’s message waits for a person to act on it, which gives the team a window to catch up.
Agents close that window. Even with solid email security in place, the occasional convincing phishing email still lands in an inbox. A person might open it seconds, minutes or hours later. An agent reads it on arrival, interprets the instructions and acts, often before anyone on the team knows the message exists.
By the time an alert fires or an analyst reviews the verdict, the agent may have downloaded the file, shared the data, changed the account or queued the payment. Faster response helps, and so does tighter automation, but anything that runs after the agent reads the message arrives second. With an agent, the action is the damage.
Detection still matters, and it has to run ahead of the agent’s first move. Your existing response process, automated and human, then serves as the backstop for what slips through.
Preemptive security flips the model of detection and response. Instead of waiting for an attack and racing to contain it, organizations anticipate threats, remove exposure and neutralize malicious content before it reaches its target, human or machine. In practice, that means five shifts:
If agents have access to email, the inbox is now an input to your business systems. Malicious content must be stopped before a person or an agent acts on it, which requires defenses that go beyond recognizing known bad links and attachments and actually understand intent and context.
Attackers research their targets before they strike, and defenders should, too. Map what an outsider can learn about your people, vendors and operations, then use those findings to harden defenses and prepare employees before a real campaign begins.
An agent that can read email and act on behalf of the inbox’s owner, with all the same permissions, is a new user with real power. Give agents only the access their tasks require, monitor what they do, and keep records that show clearly who, or what, did what.
In many organizations, IT approves new agents with little input from security. But deciding what an agent can touch is a risk decision, and it belongs with the people whose job is to anticipate abuse.
Phishing simulations and red-team exercises should include attacks aimed at agents, not only at employees. Confidence in a plan is not proof that it works.
None of this means abandoning detection and response altogether. Some threats will still get through, and teams need the ability to investigate and respond in a timely manner. But detection can no longer be the primary strategy. It has to be the backstop.
AI agents will keep spreading throughout the enterprise because the productivity gains are too enticing. Attackers know this, and they’re already testing how easily those agents can be turned against the organizations that deploy them. In the age of agentic AI, the advantage goes to whoever gets the first step, and defenders can’t afford to fall behind.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
