The Rise Of Autonomous Cyber Agents

Forbes

20-06-2025

Ronen Cojocaru, Co-CEO and Co-founder, Imperative Inc. getty
Artificial intelligence is rapidly evolving from passive tools into autonomous "agentic" systems capable of making decisions and taking actions without direct human input. These AI agents are already proving valuable as co-pilots to human analysts, enhancing threat detection and speeding up incident response.
Yet their growing autonomy is a double-edged sword. As these agents gain more power, ensuring they remain secure, transparent and reliable becomes paramount. Early examples of agentic AI in cybersecurity, from automated threat-hunting bots to self-driving network monitors, demonstrate huge potential.
However, they also highlight new vulnerabilities. AI agents can, unfortunately, be easily tricked or influenced by bad data, sometimes resorting to biased or incorrect assumptions, and users may place misplaced confidence in their outputs. In short, agentic AI is a force multiplier for cyber defense, but without proper safeguards, it can just as easily multiply cyber risk.
Despite the promise, security leaders must grapple with several emerging risks from agentic AI systems. Notably, model drift, malicious manipulation and operational reliability issues are front and center: Model Drift
Over time, AI models can become misaligned with reality as their input data changes—a phenomenon known as 'data drift.' This natural degradation in data characteristics means an AI that once performed well might start making errors as its environment evolves.
For example, an intrusion detection model trained on last year's network traffic may gradually falter as new apps, devices and attacker techniques appear. Such drift opens up new attack surfaces if not caught and corrected, undermining the model's effectiveness. Recognizing this, recent joint security guidance from the U.S. and allies urges companies to monitor AI performance closely and treat drift as an expected challenge.
Agentic AIs are vulnerable to adversarial exploits. Hackers can attempt to manipulate an AI's inputs or training data to distort its behavior. Tactics like data poisoning and feeding incorrect or malicious data into an AI's training pipeline can wreak havoc on its decision making. Imagine an attacker subtly corrupting the data that trains a spam filter or fraud detector—the AI might then start letting threats slip through or flagging the wrong items. Officials worldwide are increasingly fearful of hackers manipulating AI systems, especially those deployed in critical infrastructure. A poisoned or manipulated model not only makes bad choices; it erodes confidence that AI outputs can be trusted at all. Operational Reliability And Trust
Like all AI, autonomous agents suffer from issues of hallucination, bias and erratic behavior, which can be amplified by their autonomy. Without proper governance, an AI agent might confidently produce incorrect analyses or take unauthorized steps. These problems aren't just theoretical—early deployments have shown that AI assistants can 'go rogue' or output toxic content if misused. Businesses have learned that an unsupervised agent's mistake can lead to serious harm, reputational damage or compliance violations.
Moreover, when AI agents act unpredictably, humans tend to either over-trust them or distrust them entirely—both scenarios are risky. As one expert noted, current AI agents are still 'easily tricked' and prone to biased assumptions, yet people often trust their answers when they shouldn't. Ensuring reliability means building in rigorous testing, guardrails and oversight for AI decisions. In practice, companies are putting 'human in the loop' controls on critical uses and instituting AI red-team exercises to probe for failure modes. The goal is an AI that operates responsibly and transparently, earning trust through consistent and correct performance. Future Outlook: Roadmap For AI-Powered Cybersecurity
While today's agentic AI is still maturing, the coming years promise a dramatic expansion of AI's role in cybersecurity. In this phase, organizations move from experimentation to real deployments of agentic AI for security. AI co-pilots become common in security operations centers, handling routine tasks and assisting human analysts. For instance, autonomous AI agents might triage alerts, scour logs for threats or automate responses to basic incidents. These early agentic systems are generally narrow in scope and operate under human supervision, reflecting lessons learned about governance.
Shadow AI agents (unsanctioned bots running without oversight) emerge as a concern, prompting companies to institute AI governance programs. Industry experts emphasize the need for visibility into all AI agents in use and strict alignment with security policies to avoid 'rogue' deployments. Notably, businesses begin to treat AI agents much like employees: vetting their 'credentials,' monitoring their activities and granting only least-privilege access. As one analysis put it, AI agents can indeed augment overworked cyber teams, but only if we ensure these agents are deployed in a secure, explainable and reliable manner.
Looking a bit further out, 2026 is expected to usher in swarm intelligence and collective defense enabled by networks of AI agents. Rather than working in isolation, multiple AI systems will increasingly communicate, collaborate and even negotiate with each other across networks. Cyber defenses could be handled by fleets of specialized AI agents, with one set watching network traffic, another analyzing user behaviors and others managing endpoint security—all sharing intelligence in real time.
This coordinated 'swarm' of AI agents can respond to threats faster than any single system, mimicking a colony of ants or bees that collectively defend their nest. A new challenge will be understanding the emergent behavior of interacting AIs. When dozens of semi-autonomous agents interconnect, unexpected dynamics may arise not unlike complex financial markets or ecosystems.
By the late 2020s, the industry anticipates a transition from narrow AI tools to cognitive cybersecurity ecosystems. In practice, this means AI systems with advanced reasoning capabilities are deeply integrated into every facet of cyber defense. For example, cyber defense systems will leverage AI that emulates human-like thinking and learning processes. These cognitive SOCs can ingest vast, diverse data streams, network logs, threat intel feeds, user activity and more to make connections that human analysts might miss. Cybersecurity ecosystems will become adaptive and self-optimizing. AI will not just react to attacks but continuously learn from them, evolving its defenses.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?