AI agents at work, old attacks in overdrive

Hey there! AI has really made its mark in the daily operations of small and mid-size businesses (SMBs). It’s not just about chatbots anymore – businesses are now entrusting tasks to AI agents in order to stay competitive and level the playing field with larger companies. Some are even setting up multi-agent ‘assembly lines,’ where one supervisor agent oversees a team of specialist agents and delegates tasks between them.

But with the integration of AI comes new cybersecurity risks. SMBs often operate on tight resources, especially in the IT department. Cybersecurity responsibilities are usually handled by a small team that juggles various tasks like account provisioning and dealing with security breaches.

Despite the risks, businesses are reluctant to step back from AI adoption until all potential threats are identified and addressed. However, a recent survey found that 40 percent of SMBs don’t even have an AI policy in place. It’s a common trend for companies to implement governance only after experiencing a security incident.

Today’s security landscape is evolving in two main directions: AI opens up new avenues to business systems, while adversaries leverage AI to enhance ‘old’ threats.

AI agents, double agents, and errant agents

Every AI agent connected to a business system operates based on permissions granted by its ‘owner.’ This means that any influence on the agent’s instructions can impact how it utilizes its permissions. In a multi-agent setup, manipulated outputs can lead to cascading issues that are challenging to trace back to the root cause.

One of the risks associated with AI agents is the compromise of their supply chains and abuse of permissions. Skills, which are sets of instructions for agents, can be manipulated to carry out malicious actions, such as data theft or remote code execution.

figure-1-self-modifying-skill
Figure 1. Instructions of a self-modifying skill that can lead to unpredictable behavior and abuse (source: ESET Threat Report H1 2026)

The proliferation of poorly governed supply chains around AI agents highlights the need for stricter oversight. Continuous monitoring of skills and tool connections is essential to prevent malicious actions.

AI models, particularly large language models (LLMs), are susceptible to hallucinations that can be exploited by attackers to gain control over resources. Agents can also exhibit misalignment, leading to unauthorized actions or data breaches.

Ensuring the security of AI agents requires a comprehensive approach that considers the agent’s identity, the information it processes, the tools at its disposal, and the controls governing its actions.

Breathing new life into old attacks

While AI-specific threats are on the rise, traditional attack vectors like phishing, software vulnerabilities, and stolen credentials remain prevalent. Phishing and exploitation of known vulnerabilities are cited as the top causes of breaches in SMBs.

Attackers are constantly evolving their tactics to evade detection, with AI-powered phishing campaigns and new techniques like QR code phishing gaining popularity. Ransomware attacks are increasing in volume, targeting SMBs and leveraging tools to bypass endpoint detection and response systems.

AI is also being leveraged in malware, although current examples are still in the experimental stage. The main impact of AI in cybercrime lies in accelerating human-led attacks rather than autonomous malware development.

Where this leaves a small team

For small IT teams, navigating the complex cybersecurity landscape poses significant challenges. Hiring additional staff to address security needs may not be feasible, given the talent shortage in the cybersecurity industry.

Visibility is key to mitigating risks associated with AI agents and services. Understanding the permissions and actions of AI entities is crucial in reducing vulnerabilities. Implementing boundaries and oversight for autonomous systems can help prevent unauthorized actions.

Managed detection and response (MDR) services offer a solution for SMBs to enhance their security posture without the need for building an in-house security operations team. These services leverage AI-driven capabilities to monitor AI agents and detect suspicious activities.

By integrating advanced security measures and expert oversight, SMBs can effectively manage cybersecurity risks posed by AI adoption.

Leave a Reply

Your email address will not be published. Required fields are marked *