How AI agents are reshaping the threat landscape

The agentic AI ecosystem, powered by large language models (LLMs), is creating a new class of cybersecurity risks according to a new report.

The study from Radware finds AI agents can act autonomously, access tools and private resources, and interoperate between one another. As enterprises turn to AI agents, there is a need to govern and secure this new emerging layer of digital infrastructure.

Organizations are deploying LLM-powered agents that can reason, invoke tools, and communicate with other agents using emerging protocols. These systems are forming transitive chains of authoritative access to enterprise resources and systems that are difficult to monitor and cannot be secured with traditional protections.

Pascal Geenens. director of cyber threat intelligence at Radware, notes in the report, “We are not entering an AI future; we are already living in it. AI is no longer just a tool; it is a participant in systems, a co-author of code, a decision-maker, and increasingly, an adversary. Business leaders, security architects, and policymakers must adapt to this new reality. The agent economy presents an opportunity no business can afford to overlook. However, success will hinge on implementing it securely, as the risks are not hypothetical. The businesses that thrive will be those capable of delivering a safe, trustworthy agentic experience for their customers.”

Model Context Protocol (MCP) and Agent-to-Agent (A2A) enable agents to interact with corporate systems and data, and collaborate with other agent systems. These capabilities introduce new risks and attack surfaces such as prompt injection, tool poisoning and lateral compromise.

Malicious AI platforms are lowering the barrier for cybercrime too, subscription-based tools like XanthoroxAI offer full attack kill chain tooling that allow emerging and experienced actors to improve and automate their attacks.

LLMs can also shorten the window between responsible disclosure and exploit. GPT-4 is able to generate working exploits based on vulnerability descriptions faster than experienced security researchers. This shortens the window for defenders to respond to newly disclosed vulnerabilities and increases the risk of wide-scale attacks.

You can register for a webinar to discuss the findings on September 25th at 11am ET.

Image credit: Napong Rattanaraktiya/Dreamstime.com